Sie sind auf Seite 1von 144

Validation and Application of a Fully

Nonlinear Numerical Wave Tank


Thesis

N. Bruinsma

Author contact information

Delft University of Technology


Niek Bruinsma
Phone: +31 6 54903169
Email: niek.bruinsma@gmail.com

Delft University of Technology

Validation and Application of a Fully Nonlinear Numerical Wave Tank


VALIDATION AND A PPLICATION OF A
F ULLY N ONLINEAR N UMERICAL WAVE
TANK
T HESIS

by

N. Bruinsma

Master of Science

in Offshore and Dredging Engineering


at the Delft University of Technology.

Student: Niek Bruinsma


University supervisor: Prof. R. H. M. Huijsmans
Dr. ir. I. Akkerman
Dr. ir. M. Zijlema
Dr. ir. G. H. Keetels
Company supervisor: Dr. ir. B. T. Paulsen
P REFACE
The present master thesis is submitted as part of the requirement for obtaining a master of
science degree in Offshore Engineering at the Technical University of Delft. The work has
been conducted at Deltares, within the department of Harbour, Coastal and Offshore En-
gineering, under supervision of Dr. ir. Bo T. Paulsen. Supervision and guidance from the
Technical University of Delft was provided by Prof. René H. M. Huijsmans, Dr. ir. Ido Akker-
man, Dr. ir. Marcel Zijlema and Dr. ir. Geert H. Keetels. This thesis was performed as part
of the ’TO2-Floating Wind’ project, and the success of the presented work was stimulated by
the collaboration of ECN, MARIN, NLR and Deltares.
I would like to express my gratitude towards Deltares for hosting me during my research,
and MARIN for providing the experimental data essential for the success of the of this work.
I am grateful for the guidance and supervision from my supervisors from the Technical Uni-
versity of Delft, but in particular for the unlimited support provided by Dr. ir. Bo T. Paulsen,
who, to a great extent, is responsible for the success of this work. Finally, I would also like
to express my gratitude towards Dr. ir. Niels G. Jacobsen for his support and stimulating
discussions throughout the project.

N. Bruinsma
Delft, March 2016

iii
A BSTRACT
The present thesis considers numerical computations of fully nonlinear fluid-structure inter-
action. The aim of the thesis is to establish a well validated fully nonlinear numerical wave
tank for the simulations of complex fluid-structure interaction of moored floating offshore
structures.
The numerical computations are carried out using the fully nonlinear numerical fluid-
structure interaction solvers, interFoam and interDyMFoam, from the open-source CFD li-
brary OpenFOAM®. These solvers make use of a fully nonlinear Navier-Stokes/VOF solver for
the computations of the two-phase flow field, where the interDyMFoam solver also utilises a
6-DOF motion solver to compute the motions of the floating structure. These solvers were ex-
tended with waves2Foam, a wave generation and absorption toolbox, developed by Jacobsen
et al. (2012). The extended interFoam and interDyMFoam solvers, hereafter referred to as the
waveFoam and interDyMFoam solver, were utilised for the computations of fluid-structure
with fixed and moving structures respectively. Furthermore, an implementation of catenary
mooring lines is provided by Niels Jacobsen (Researcher/advisor at Deltares), for the simu-
lations of the mooring system of the floating structure. Finally, the fully nonlinear potential
flow solver, OceanWave3D, was utilised in a fully nonlinear and fully parallelised domain de-
composed solver, developed by Paulsen (2013) and Paulsen et al. (2014b), for the efficient
computations of realistic sea states. Here, the outer wave field is described by the potential
flow solver, whereas the inner wave field, in the vicinity of a given structure, is described by
the interDyMFoam solver.
The waveFoam and waveDyMFoam solvers are carefully validated either in terms of con-
vergence by grid refinement or by comparisons to experimental measurements. Special at-
tention is paid to the waveDyMFoam solver with respect to the computation of the flow-
induced motions of a moored floating wind turbine. The ability of the numerical model to
accurately reproduce experiments, performed with the generic OC5 floating wind turbine
model in the MARIN Concept Basin, is also investigated.
The Navier-Stokes/VOF based waveFoam and waveDyMFoam solvers were evaluated with
respect to wave propagation and wave structure interaction in a two-dimensional set-up. A
grid convergence study was performed on the propagation of a fully nonlinear stream func-
tion wave. The convergence rate of the two-phase numerical solution to the single-phase
stream function solution was verified. Successful computation of wave loading on a partially
submerged fixed horizontal cylinder was provided. The accuracy of the numerical model,
with respect to fluid-structure interaction for free and forced motion of a structure, was ver-
ified with the generation of surface waves by the forced oscillation and the free heave decay
of a horizontal cylinder.
These two-dimensional cases were also used to verify the efficiency of two meshing tools,
provided by the OpenFOAM® toolbox. These meshing tools were shown to provide excellent
results, and more importantly, provided a less complicated method for generating surface
boundary meshes of complex three-dimensional structures.
The waveDyMFoam solver is used for the computation of free and moored decay test of

v
vi A BSTRACT

the three-dimensional floating wind turbine model. The accuracy of the numerical solution
was verified against numerical computations from the work of Dunbar et al. (2015) and phys-
ical experiments performed at MARIN.
Finally a proof-of-concept case is performed, involving the modelling of a three-dimen-
sional moored floating wind turbine subjected to irregular uni-directional waves. In these
numerical computations the potential of the waveDyMFoam and the domain decomposition
strategy are evaluated.
C ONTENTS

1 Introduction 1
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Wind energy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Floating wind energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Design tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.4 International collaboration . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.5 National collaboration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.6 Hydrodynamical tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Free surface waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Wave parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.2 Small linear waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Steep nonlinear waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.2.4 Wave loads on a surface piercing cylinders . . . . . . . . . . . . . . . . . . 8
1.2.5 Wave interaction with a floating body. . . . . . . . . . . . . . . . . . . . . 10
1.3 Numerical simulation of wave-structure interaction . . . . . . . . . . . . . . . . 11
1.3.1 Flow theories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.2 Navier-Stokes-type models . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.3 Navier-Stokes/VOF solvers. . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.4 Efficient domain decomposition . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 This Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.5 Structure of this thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 The numerical models 17


2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 InterFoam/waveFoam solver . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2.1 Governing equations Navier-Stokes/VOF. . . . . . . . . . . . . . . . . . . 18
2.2.2 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2.3 Relaxation zones . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.2.4 Algorithm waveFoam. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3 InterDyMFoam/waveDyMFoam solver. . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.1 Motions of a body . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3.2 Under-relaxation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.3 Dynamic mesh handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.4 Rotations and translations . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.5 Restraints for moving boundary surfaces . . . . . . . . . . . . . . . . . . . 26
2.3.6 Algorithm waveDyMFoam . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4 Fully nonlinear potential flow solver . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.4.1 Governing equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4.2 Algorithm waveDyMFoam / potential flow . . . . . . . . . . . . . . . . . . 31

vii
viii C ONTENTS

3 Wave propagation and fluid-structure interaction in 2D 33


3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2 Convergence study on the propagation of a fully nonlinear regular wave . . . . . 33
3.2.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.2 Analysis convergence study . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Wave loading on a partially submerged fixed horizontal cylinder . . . . . . . . . 40
3.3.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3.2 Analysis wave loading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Surface waves generated by the forced oscillation of a horizontal cylinder . . . . 47
3.4.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.4.2 Analysis wave generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.5 Free heave decay of a horizontal cylinder . . . . . . . . . . . . . . . . . . . . . . 53
3.5.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.5.2 Analysis motion decay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

4 Meshing tools 61
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
4.2 SnappyHexMesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.2.1 Wave loading on a partially submerged fixed horizontal cylinder. . . . . . 63
4.2.2 Surface waves generated by the forced oscillation of a horizontal cylin-
der . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2.3 Free heave decay of a horizontal cylinder. . . . . . . . . . . . . . . . . . . 65
4.3 Multi-mesh grading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
4.3.1 Wave loading on a horizontally fixed cylinder . . . . . . . . . . . . . . . . 67
4.3.2 Surface waves generated by a forced heaving cylinder. . . . . . . . . . . . 68
4.3.3 Decay of heave motion of a free horizontal cylinder . . . . . . . . . . . . . 68
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

5 Fluid-structure interaction in 3D 71
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.2 TO2 floating wind project . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.3 Physical experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3.1 Physical wave tank . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3.2 OC5 floating offshore wind turbine model . . . . . . . . . . . . . . . . . . 73
5.3.3 Experiments performed at MARIN . . . . . . . . . . . . . . . . . . . . . . 73
5.4 Free heave, pitch and roll decay of a floating structure . . . . . . . . . . . . . . . 74
5.4.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.4.2 Analysis of the free decay simulations . . . . . . . . . . . . . . . . . . . . 77
5.5 Motion decay of a floating structure with restraints . . . . . . . . . . . . . . . . . 81
5.5.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.5.2 Analysis of the moored decay simulations . . . . . . . . . . . . . . . . . . 83
5.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

6 Proof-of-concept 89
6.1 Wave propagation using OceanWave3D . . . . . . . . . . . . . . . . . . . . . . . 89
6.1.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
6.1.2 Analysis of wave propagation using OceanWave3D . . . . . . . . . . . . . 91
C ONTENTS ix

6.2 Moored floating structure subjected to irregular waves . . . . . . . . . . . . . . . 94


6.2.1 Numerical set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.2.2 Analysis moored floating structure subjected to irregular waves . . . . . . 95
6.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

7 Conclusions and recommendations 103


7.1 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
7.1.1 A converging solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7.1.2 Propagation of a fully nonlinear wave. . . . . . . . . . . . . . . . . . . . . 104
7.1.3 Wave loads on a structure . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.1.4 Motions of a structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.1.5 Advanced meshing tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
7.1.6 Fully nonlinear numerical wave tank . . . . . . . . . . . . . . . . . . . . . 107
7.2 Conclusions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.3 Further developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
7.4 Future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

A WaveFoam code 111

B WaveDyMFoam code 115

C SixDoFRigidBodyMotion code 119

D Floating wind turbine model 127

Bibliography 131
1
I NTRODUCTION

1.1. B ACKGROUND
Scientific understanding of global warming and the major environmental risk it poses is in-
creasing and the notion that global warming is caused by increasing concentrations of green-
house gases and other human activities has been widely recognized. Major industrialised
nations are working together on possible societal responses, the most apparent being miti-
gation by emissions reduction. Part of this mitigation strategy is to change the worlds energy
supply from being predominately dependent on fossil fuels to the use of low-carbon energy
technology such as nuclear-, solar-, thermal-, hydro- and wind energy.
Recently the leaders of the European Union (EU) agreed on "The 2020 Climate and Energy
Package", a set of binding legislation which aims to ensure the EU meets its ambitious climate
and energy targets for 2020. These targets set three key objectives for 2020: a 20 % reduction in
greenhouse gas emissions compared to the 1990 levels, 20 % energy consumption produced
from renewable resources and a 20 % improvement in the EU’s energy efficiency (Moccia
et al., 2014). It may be noted that the environmental risk is not the only reason for the EU’s
increased interest in alternative energy sources. In recent years the relationship with Russia
seems to deteriorate, resulting in an accumulated desire to be less dependent on Russian oil
and gas. All this has stimulated investments in alternative energy sources.

1.1.1. W IND ENERGY


It is expected that wind energy will be a major contributor to the increase of renewable energy
production as this technology is highly suitable for wide implementation throughout the EU.
According to the European Wind Energy Association (EWEA) the total installed wind energy
capacity in the EU amounted to 117.3 GW by the end of 2013. The majority of this capacity,
110.7 GW, was contributed by onshore wind farms and the remaining 6.6 GW came from
offshore wind farms. That same year, 32 % of all the newly installed capacity was attributed
to wind energy. This made wind power the fastest growing energy technology in terms of
capacity growth. The wind energy capacity installed by the end of 2013 would, in a normal
wind year produce 257 TWh of electricity, which is enough to cover 8 % of the electricity
consumption of the EU, this is an increase of 1 % compared to 2012. EWEA predicts the
combined wind energy production (offshore and onshore) to cover 14.9 % of the total EU
power demand. This increased capacity will to some extent be achieved by replacing active

1
2 1. I NTRODUCTION

Figure 1.1: Schematic representation of different types of offshore wind support structures, the designs of which
are adapted from the offshore oil and gas industry. From left to right: monopile, jacket, semi-submersible, tension
leg platform and spar buoy.

wind turbines by newer, larger and more efficient models, but for the greater part by the
installation of new wind farms (Pineda et al., 2014).
Offshore wind energy installations are expected to produce approximately 2.9 % of the
total EU energy demand by 2020. This projected growth of offshore wind energy corresponds
to the growth witnessed in the onshore wind sector at a similar time in that industry’s de-
velopment. Onshore wind energy deployment picked up speed in the mid-1990s. With a
15 year difference, presently, the offshore wind seems to be following a similar growth path.
The foreseen growth of the offshore wind sector will move it from an emerging, immature
technology to a key component of the EU’s energy mix (Arapogianni et al., 2011). As the tech-
nology develops, the offshore wind industry is moving into deeper waters, further from the
coast with bigger farms. This reflects maritime spatial planning and wind farm developers
desire to harness better energy resources further out at sea.

1.1.2. F LOATING WIND ENERGY


Although currently ongoing commercial scale offshore wind developments are based on con-
cepts with fixed substructures, many countries have limited suitable sites in sufficiently shal-
low water to economically install wind turbines on monopile and jacket foundations, see Fig-
ure 1.1. A floating offshore wind turbine (FOWT) concept was introduced to harvest the vast
energy resources found in the deeper waters. This concept has existed since the early 1970s,
however, first full size test cases only initiated in 2008 (Arapogianni et al., 2013). The most
apparent advantage of floating offshore wind over conventional offshore wind, is the possi-
bility access the vast wind resources, located at deeper sites further out at see. Furthermore,
since a FOWT will generally be located further from the coasts, the visual and noise impact
to inhabited coastal areas will also be lower. On top of that, the relatively simple installa-
tion process is another major upside to this technology. The installation of the conventional
offshore wind relies heavily on expensive special purpose ships for transport and on-site as-
sembly and installation, while a FOWT can be assembled in port and towed into position
using simple tugboats, whereupon the mooring system and electric grid can be connected.
Currently several FOWT concepts are actively undergoing full-scale prototype testing.
The "Hywind", developed by Statoil, was installed in the North Sea, just off the Norwegian
coast and has been running successfully since 2009. The spar type structure with a 2.3 MW
1.1. B ACKGROUND 3

turbine was towed 10 km offshore, where it was installed in 220 m deep water. Statoil plans
to build the first floating wind farm off the Scottish coast. The 30 MW pilot project, which
should be commissioned by the end of 2017, will consist of five, 6 MW floating wind tur-
bines operating in waters exceeding 100 m of water depth. The "WindFloat", developed by
the join-venture under the name of Windplus, was installed off the coast of North Portugal
in 2011. This semi-submersible tri-floater has been successfully tested in water depths of ap-
proximately 45 m. Based on their sub-structure and mooring system, these, and other equiv-
alent, FOWT concepts can be categorised into three main designs: the semi-submersible, the
tension leg platform and the spar buoy. These designs, presented in Figure 1.1, were adapted
from the offshore oil and gas industry.

1.1.3. D ESIGN TOOLS


Offshore floating wind is seen as a promising solution to harvest the deep offshore wind re-
sources. However booming, the offshore wind sector is constrained by high cost and rela-
tively small economical margins compared to the offshore oil and gas industry. Reliable and
purpose oriented design methods are essential to make this technology more cost efficient.
Some of the current engineering tools that are used to design offshore wind systems are
based on design tools developed for land-based wind turbines. These design tools use codes
that incorporate aerodynamic, control system, and structural-dynamic models. However,
because offshore wind turbines are installed in marine environments, they require modelling
tools that can also consider the hydrodynamic loads on a variety of substructure types as
well as their mooring systems. To design these systems, engineers need engineering tools
that can simulate incident waves, sea current, hydrodynamics, dynamics of the substructure
and mooring system. The complex nature of these tools emphasises the need to verify and
validate their accuracy.

1.1.4. I NTERNATIONAL COLLABORATION


The Offshore Code Comparison (OC3) project, part of the International Energy Agency (IEA)
Wind Task 23, 2005-2009, was the first in a series of international collaboration projects that
focused on the validation of offshore wind modelling tools (Jonkman and Musial, 2010). The
two-phased Offshore Code Comparison Collaboration Continuation (OC4) project, IEA Wind
Task 30, 2010-2013, was a continuation of the OC3 project (Jonkman et al., 2012). This IEA
Wind Task 30 is currently extended with the Offshore Code Comparison Collaboration, Con-
tinued, with Correlation (OC5). These projects aim to verify the accuracy of complex en-
gineering tools, utilised for the design of offshore wind systems, through code-to-code and
code-to-data comparisons.
The Offshore Code Comparison Collaboration projects involve the modelling of fixed and
floating wind systems. The first phase of the OC4 project involved the testing of three differ-
ent generic floating wind turbine designs in a wave basin at Marine Research Institute Nether-
lands (MARIN) (Koo et al., 2014). To cover the main design concepts for the offshore floating
wind industry these three design concepts included a spar, a tension leg platform and a semi-
submersible, see Figure 1.1. The second phase of the OC4 project and the currently ongoing
OC5 project concentrated on further development and validation of design codes using the
semi-submersible FOWT.
The most resent semi-submersible FOWT design, the OC5, presented in Figure 1.2, fea-
tures a 5MW horizontal-axis wind turbine with pitch controlled blades, developed by the Na-
4 1. I NTRODUCTION

Figure 1.2: Schematic representation of the OC5 FOWT, a semi-submersible floating wind system consisting of
a 5MW horizontal-axis wind turbine with pitch controlled blades mounted onto a semi-submersible try-floater,
which is moored to the seabed using three catenary mooring lines.

tional Renewable Energy Laboratory (NREL) (Jonkman et al., 2009). This turbine is mounted
onto the semi-submersible try-floater and moored to the seabed with spread three lines cate-
nary mooring system. It may be noted that the generic nature of all the FOWT floater designs
developed for or by the Offshore Code Comparison Collaboration projects implies that they
are publicly available.

1.1.5. N ATIONAL COLLABORATION


The OC5 FOWT design was also utilised for a joint research project aimed at increasing the
knowledge on floating offshore wind turbines. During 2015, four Dutch research institutes,
ECN, MARIN, NLR and Deltares, worked together on this government subsidised project, re-
ferred to as the ’TO2-Floating Wind’ project. A more detailed description of the project as
well as the model can be found in Chapter 5.
The present research was conducted as part of the contribution from Deltares to this TO2-
Floating Wind project. The main focus of Deltares was to extend the knowledge concerning
numerical simulation of the motion of floating wind turbines and the calculation of loads in
the mooring system. The project strived for code-to-data analysis, which was partly accom-
modated by physical model tests performed at MARIN.

1.1.6. H YDRODYNAMICAL TOOLS


At the moment no large body of knowledge exists of the behaviour of floating wind systems,
hence there is still much work to be undertaken to improve their performance. Modelling
such a system, especially in cases where the flow is more violent, can help to understand its
behaviour in real seas.
Conventional hydrodynamical engineering tools, as used in the oil and gas industry, are
often developed for designing floating structures which are installed in water depths of over
300 m. At these larger water depths linear wave theory may be sufficient to describe wave
profiles and determine wave loading. However, the offshore wind industry initially intends
1.2. F REE SURFACE WAVES 5

Figure 1.3: Schematic representation of types of waves which can cause loading impacts on a floating wind tur-
bine. (a) Linear waves, (b) Steep nonlinear waves and (c) highly nonlinear breaking waves.

to install floating structures at intermediate water depths, from 50 m to 200 m. Here, the
waves can become strongly nonlinear in case of extreme weather conditions. Utilising linear
or second-order wave theory is not sufficient to provide a realistic description of the wave
profile and lack accuracy in terms of the determination of extreme wave loads.
Physical model testing is at present the only reliable method for investigating the strongly
nonlinear hydrodynamical flow effects that occur during these extreme weather events. Con-
structing and testing a model is expensive, let alone conducting an intensive model campaign
with multiple models subjected to a variety of wave conditions. It is often not possible, or
highly expensive, to make adjustments to the model or the measurement equipment during
these experiments and the number of sea states that can be tested is restricted by the time
available in the test basin, thus of high influence to the total cost of the model campaign.
Furthermore, only local measurements of the surface elevation, flow velocities and pressures
can be acquired since they depend on the physically installed measurement equipment.
Computational Fluid Dynamics (CFD), which has been used as a successful design tool in
many other many other areas of engineering, could be an alternative approach for modelling
of a FOWT. It is a numerical method for the calculation of the nonlinear differential equations
describing a fluid flow. One of the major benefits of numerical modelling over scale model
experiments are the flexibility in the set-up of design cases, last minute design changes or
changes to the environmental conditions. Also, since the modelling takes place in a numer-
ical domain, it is possible to take measurements of the free surface, flow velocities and pres-
sures at any given point in space and time. Furthermore, by using nonlinear equations for
describing the air and water flow, CFD allows for the application of a much wider range of
wave models, that can actually model higher order nonlinear effects. However, in order to
use CFD as a reliable design tool for floating offshore structures, the accuracy of predictions,
numerical convergence and computational efficiency need to be validated.

1.2. F REE SURFACE WAVES


To get a better understanding of the processes involved in fluid-structure interaction, this
section will address different types of free surface waves, their measure of nonlinearity, the
possible load effects they can have on a structure and the available wave theories that can
be used to describe them. In the present work, waves are categorised as small linear waves,
steep nonlinear waves and highly nonlinear breaking waves. A schematic representation of
these three types of waves is presented in Figure 1.3.
6 1. I NTRODUCTION

Figure 1.4: Schematic representation of a steep wave in a regular wave train (Fenton, 1990). Here, the wave length
is denoted by λ and the wave celerity, c, is defined as c = Tλ , where the wave period, T , is the time between
consecutive crests.

1.2.1. WAVE PARAMETERS


A wave train is a series of waves, like the one presented in Figure 1.4, travelling in the same di-
rection, spaced at regular intervals. There are three physical dimensions that uniquely define
a wave train: the water depth d , the wave height H , which is delimited crest-to-trough, and
the wavelength λ. There are many wave theories where the wave period, T , the time between
two consecutive crests, replaces the wave length as the third parameter. The wavelength and
wave period are affiliated through the wave velocity relative to the current (Fenton, 1990).
This relative wave velocity is often referred to as the wave celerity and can be defined as c = Tλ .
Another important parameter of a wave is its steepness, which can be described by the ratio
of the wave amplitude, A, and the wave length, λ. The wave steepness is a measure for the
nonlinearity of the wave. A regular wave also has a limiting theoretical wave height, Hmax ,
which is dependent on the water depth and the wave length. According to the theoretical
work of Williams (1981) and Fenton (1990), this limiting wave height for regular waves can be
defined by

Hmax 0.141063 dλ + 0.0095721( dλ )2 + 0.0077829( dλ )3


= . (1.1)
d 1 + 0.0788340 dλ + 0.0317567( dλ )2 + 0.0093407( dλ )3

Here, Hmax can be used as an alternative method for describing the measure of wave non-
linearity. The ratio of the wave height, H , and the limiting wave height, Hmax , gives a more
consistent measure for the nonlinearity across various water depths than the ratio of the wave
amplitude and the wave length.

1.2.2. S MALL LINEAR WAVES


Small non steep regular waves, travelling in intermediate to deep water, where the free sur-
face elevation is described by a simple sine function, are in general well defined by linear
wave theory. The validity of various wave theories has been thoroughly investigated through-
out the last decades, the result of which is presented in Figure 1.5. Here, the validity of a wave
theory is presented as a function of the wave height and the water depth. It may be noted that
these analytical wave theories are only valid upto the theoretical breaking limit, Hb . It can be
observed that the wave height for which linear theory gives a valid description of the regular
waves declines with the water depth. This can be explained by the fact that the profile of a
wave changes as the wave moves towards shallower water. During this transition the wave
1.2. F REE SURFACE WAVES 7

Figure 1.5: Validity of various wave theories presented as a function of the wave height and the water depth, both
of which were normalized by the gravitational acceleration multiplied by the wave period squared. Hb is the
theoretical breaking limit (Det Norske Veritas, 2014).

crest tends to get shorter while the trough gets flatter, i.e. the measure of nonlinearity of the
wave increases. These changes to the wave profile are an effect of the enlarged influence of
the seabed. This phenomenon is considered to take effect from the deep water limit, kh ≈ 3,
where kh is the non-dimensional water depth, which is the water depth multiplied by the
wave number, k = 2π λ .
Small non steep waves may be significant to the Fatigue Limit State (FLS) of an offshore
structure due to their abundance in real sea states. However, their immediate impact load,
when compared to extreme waves during storm events, is relatively insignificant to the Ul-
timate Limit State (ULS) of an offshore structure and therefore considered to be outside the
scope of the present research.

1.2.3. S TEEP NONLINEAR WAVES


Waves closer to the theoretical breaking limit, Hb , especially in shallow and intermediate
water depths, are considered to have significant steepness. This implies that they are char-
acterised by shorter crests and flatter troughs and do not resemble a single sine function, see
Figure 1.4 for a schematic representation of such a wave profile. Therefore, linear wave the-
ory can no longer be used to give an accurate description of the wave profiles of such waves.
Instead, higher order wave theories, e.g. higher order Stokes, Cnoidal or the nonlinear stream
function (Fenton, 1990), can be used to describe regular steep waves. The validity of various
orders of the Stokes and the stream function theories are presented in Figure 1.5. It may be
noted that these theories are restricted to the description of regular waves, or wave trains,
and a flat sea bed and so do not represent realistic irregular sea states.
A limitation to the use of both the Stokes and Cnoidal theory is that they are specific to
certain types of waves and thus not generally applicable. Stokes theory is known to lack in ac-
curacy when describing shallow water waves, while the accuracy of Cnoidal theory is limited
when it comes to the description of deep water waves. Both theories face difficulties describ-
ing extremely high waves. Various efforts were undertaken to develop methods which did
not face these problems. Most of these wave theories focussed on the numerical determina-
tion of the Fourier coefficients, instead of using perturbation expansions as was done for the
Stokes theory. This finally led to the fully nonlinear stream function method by Rienecker and
8 1. I NTRODUCTION

Figure 1.6: Relative importance of inertia, drag and diffraction to the inline forces, on vertical surface piercing
cylinders subjected to incoming regular waves, presented as a function of the wave height, H , and length, λ,
normalized by the cylinder diameter (Det Norske Veritas, 2014).

Fenton (1981), that could be utilised for the accurate description of steep nonlinear waves in
both deep water and water of finite depth. This theory was later presented in a computer
program by Fenton (1988).
Wave breaking, a process which causes large amounts of wave energy to be transformed
into turbulent kinetic energy, can occur in the proximity of the theoretical breaking limit.
Sarpkaya and Isaacson. (1981) indicated that there are four possible types of wave break-
ing: spilling, collapsing, plunging and slurging. Such overturning waves are considered to
be strongly nonlinear, where, air and water layers are mixing and multiple free surfaces start
forming. It should be noted that no analytical solutions exist for these kind of waves.

1.2.4. WAVE LOADS ON A SURFACE PIERCING CYLINDERS


The phenomenon of wave loads on vertical surface piercing circular cylinders is a much re-
searched topic since they appear as a fundamental part of many offshore structures. The
inline forces on vertical surface piercing cylinders, subjected to incoming regular waves, can
be classified as: drag, inertia and diffraction. The governing type is determined by the ra-
λ
tio between the wave length and the cylinder diameter, D , and the ratio between the wave
height and the cylinder diameter, H D . Figure 1.6 can be used as a guideline for establishing
the dominant force type in a specific design case. If a cylinder is relatively wide compared to
λ
the wavelength, D < 5, wave diffraction is considered to be the dominant load factor, while
λ
the inertia forces are dominant for D > 5 and drag forces become dominant for long waves
where the cylinder is relatively slender compared to the height of the wave, or H D > 10.
In context of the present research, the size of the main tubular members of the OC5
FOWT, D = 12m combined with the relatively high, H ≈ 20m, and long waves, λ ≈ 200m,
relevant for the ULS design of a floating offshore structure in intermediate water depth, give
λ
rise to moderate HD and a large D . This would indicate that inertia forces play a dominant role
in the investigation of these extreme wave load events. It may be noted that, not the direct
impact of a large wave, but rather a secondary phenomenon, such as spray or wave run-
up against a more fragile sub-component can be the cause of significant structural damage.
Highly detailed numerical or physical models are needed to evaluate the potential dangers of
these nonlinear fluid-structure interactions.
Analytical solutions to wave forcing by regular waves on vertical cylinders, that include
second, and third-harmonic load components, have been derived by Faltinsen et al. (1995)
and Malenica and Molin (1995). However, these solutions are only valid for regular non-steep
waves and to date no higher order analytical solutions for wave forcing exist for real irregu-
1.2. F REE SURFACE WAVES 9

lar sea states. At present the Morison equation (Morison et al., 1950) is widely applied for
the calculation of hydrodynamic loads on offshore structures. This semi-empirical equation
for determining the inline force on a body subjected to a flow consists of two parts; an iner-
tia force and a drag force. State-of-the-art offshore wind turbine design codes, like the Det
Norske Veritas (DNV), prescribe the Morison equation to calculate the inline wave forces on
slender structures, such as tubular members of jackets and monopiles.
When it comes to calculating the loads on FOWT support structures, the Morison equa-
tion has a couple of shortcomings. Three of the main disadvantages were pointed out by
Matha et al. (2011). First, to simplify the diffraction problem, the Morison equation assumes
the wave potential to be constant across the body. This is in accordance with the long-
wavelength approximation, so for bodies with a small diameter relative to the wave length.
However, for many FOWT designs the body is relatively large and the wave disturbance may
therefore not be neglected. Diffraction effects must be included if one is to correctly deter-
mine the local pressure forces and global wave loads. Secondly, the Morison equation as-
sumes that viscous drag dominates the drag loading, and that wave radiation damping can
therefore be ignored. This assumption is valid for slender structures and small motions, how-
ever, FOWT are in general of considerable volume and experience significant movement so
that radiation forces should be taken into account. Lastly, the Morison equation does not
take into account the added mass-induced coupling between the hydrodynamic force and
the support structures accelerations. Apart from the limitations pointed out by Matha et al.
(2011), the Morison equation is not able to describe strongly nonlinear free-surface effects
encountered at fluid structure interaction in extreme sea states, e.g. large wave run-up, spray
and wave breaking. Furthermore, the Morison equation fails to give detailed description of
high-frequency loading, which is presumed to be critical for describing nonlinear phenom-
ena such slamming, springing and so called ’ringing’.
Within the context of this research slamming is considered to be a hydrodynamic impact
caused by a steep or braking wave. This strongly non-linear fluid-structure interaction can
cause very large impact forces, which are highly dynamic and have a short duration. When it
comes to FOWT structures, slamming loads can cause large local damage, but also accelerate
the fatigue of the global structure due to the resulting dynamic response. The relative im-
pact angle and the relative fluid-structure velocity are considered to be the main parameters
determining slamming loads.
Springing is one of the more commonly observed non-linear phenomenon. It is a steady
state resonance phenomenon caused by second-order wave effects, see Figure 1.7 for a rep-
resentation of a springing response signal. Springing is observed in the vertical and bending
modes of tension leg platforms and gravity based structures where it affects the fatigue level
of the structure. Springing is therefore considered to be more significant to the FLS design
then to the ULS design and will therefore not be incorporated in the present research.
Ringing, a potentially more dangerous nonlinear phenomenon, was first observed in some
tension leg platforms and gravity based structures in the North Sea during the 1980’s. Since
the discovery of the phenomenon great efforts have been made to identify the wave mecha-
nisms that induce it in offshore structures. Ringing is a strong high-frequency transient re-
sponse believed to be caused by very steep waves in extreme sea states. The structural re-
sponse builds up over approximately one wave period and decays to steady state at a rate
that depends on amount of damping in the system, see Figure 1.7 for a representation of a
ringing response signal. Ringing of vertical cylindrical members of offshore structures was
10 1. I NTRODUCTION

Figure 1.7: Amplitude signal indicating a springing and a ringing event (Waisman et al., 2002).

once sought in third order theory, because early observations indicated third order wave fre-
quencies to be corresponding to the natural frequency of the structure. A substantial amount
of studies regarding the ringing phenomenon has been carried out since and it is now defi-
nitely known to be a strongly nonlinear phenomenon Rainey (2007). Higher harmonic non-
linear forcing as well as a secondary loading were identified by Grue and Huseby (2002), to be
causing these ringing vibrations at a resonance frequency.

1.2.5. WAVE INTERACTION WITH A FLOATING BODY


Linear potential theory for waves acknowledges three contributions to the free surface eleva-
tion in the surroundings of a floating structure. First, the undisturbed incoming waves, which
are present in spite of the presence of the structure. Secondly, the diffracted waves, generated
due to the reflection of wave energy against the structure. Lastly, the radiated waves, resulting
from the motion of the structure. For slender structures, waves pass relatively undisturbed,
this is why the Morison equation assumes the wave potential to be constant across the body.
However, significant disturbances to the wave kinematics, i.e. wave diffraction, are observed
when the structure is relatively large compared to the wave length.
The DNV design code (Det Norske Veritas, 2014), states that wave diffraction becomes
important for cylindrical shaped structures with a non-dimensional cylinder diameter kD '
1.3. Here, the non-dimensional cylinder diameter is the product of the wave number, k = 2π λ ,
and the cylinder diameter, D. This would imply that, if one considers the 12 m diameter
cylindrical members and a wave climate where the highest waves have wave lengths in the
order of 200 m, wave diffraction is not an important analysis criteria for a floating structure
like the OC5 FOWT, since the non-dimensional cylinder diameter would stay well below the
defined limit. It may however be the case that, if one looks at the structure as a whole instead
of the individual tubular members, the presence of the relatively large structure can result
in significant disturbances to the incoming waves. This can only be investigated if wave-
structure interaction is taken into account.
Many design tools that are utilised to calculate hydrodynamic loads on offshore struc-
tures, were developed for stationary bodies. These models are often based on potential-flow
theories which are only valid when the translational motion of the structure is small relative
to the wavelength and the rotational motion is less than the wave steepness. Many float-
ing configurations, however, can experience large translational displacements as a result of
the low resistance to surge and sway in their catenary moored systems. This was denoted in
Matha et al. (2011), where it was also concluded that these models, based on potential-flow
theories, are no longer valid when it comes to the calculation of hydrodynamic loads on float-
ing offshore wind turbines. Furthermore, the DNV design codes (Det Norske Veritas, 2014),
state that the large translations of a FOWT, due to incoming waves, can produce significant
1.3. N UMERICAL SIMULATION OF WAVE - STRUCTURE INTERACTION 11

radiation waves, therefore, radiation forces should be incorporated in the design analysis.
In the ULS design of a floating offshore structure, the analysis often focusses on an ex-
treme wave event, for instance the impact of a one in fifty year wave. Such a wave has the
potential to cause large damages to the the structure or the mooring system. The main struc-
tural components of the floating support structure can locally experience extremely high
forces, in the form of inertia, diffraction, radiation or drag forces, caused by the direct im-
pact of a wave. Also, weaker secondary components can be exposed to indirect loading from
wave spray or run-up. It should be noted that, at present, performing highly detailed model
experiments is the only validated approach to analyse these strongly nonlinear phenomena.
These experimental analysis could possibly be supplemented, if not substituted, by simu-
lations performed with a well validated numerical model capable of accurate simulation of
wave-structure interactions of a moored floating structure.

1.3. N UMERICAL SIMULATION OF WAVE - STRUCTURE INTERACTION


To get an idea about the potential of numerical models with respect to engineering problems
concerning free surface flow and fluid-structure interaction, this section will address some of
the most important numerical methods currently at the disposal of hydrodynamic engineers.

1.3.1. F LOW THEORIES


The numerical simulation of free surface flow problems is a challenging topic, let alone with
the presence of a moving structure, however, it has been of considerable importance to the
development of offshore structures. Several types of models are commonly seen in numerical
simulation of the non-linear waves and wave-structure interaction. These models can be
categorised as, either, Boussinesq, potential flow, or Navier-Stokes-type models, referring to
the governing equations driving the model.
Boussinesq-type models have been widely used to solve free surface flows since the com-
putational costs are reasonable and the accuracy of the approximation of water waves is valid
for linear and weakly nonlinear waves, given that the wavelength is large compared to the wa-
ter depth. A review of the developments of these types of models, as well as the description
of a state-of-the-art model, is given by Madsen and Fuhrman (2010).
Potential flow theory has been very successful in modelling the wave loading on offshore
structures and has become a standard tool in ocean engineering. The potential solver WAMIT
is based on linear and second-order potential theory and is considered one of the best of its
class, it has the capability of representing the geometry of the structure by a higher-order
method(Lee, 1995). This is just one of the industries many state-of-the-art numerical models
which are based on the potential flow theory, two other good examples are AQUAPLUS (Del-
hommeau, 1993) and AEGIR (Kring et al., 1999). These potential flow models in general well
equipped to calculate first and second order wave loads and fluid-structure interaction at ac-
ceptable computational costs. They are very useful for demonstrating the initial feasibility of
a floating structure. For instance by identifying the natural frequencies of the system, which,
in order to minimise the dynamic response, need to be placed away from the wave spectrum.
However, for the investigation of the strongly nonlinear hydrodynamics phenomena in-
volved with wave-structure interaction during extreme wave events, the potential flow the-
ory is inadequate and fully non-linear numerical methods are required. The advancements
in computational power in the last years have paved the way for fully nonlinear CFD mod-
12 1. I NTRODUCTION

els. These solve the Navier-Stokes equations in combination with the continuity equations to
compute strongly nonlinear wave loads, fluid-structure and in some models it is even possi-
ble to simulate surface effect such as wave breaking.

1.3.2. N AVIER -S TOKES - TYPE MODELS


The Navier-Stokes-type models can either be meshless or grid-based, each of these types has
its own advantages and disadvantages. A major distinction between the two is the need for
an additional equation to describe the free surface.
Smoothed-particle hydrodynamics (SPH) is a renowned fluid particle based meshless
Navier-Stokes-type method that utilises a Lagrangian specification of the flow field, i.e. the
coordinates move with the fluid. The advantages of meshless methods is that they are gen-
erally better at dealing with large local distortions than grid-based methods. A thorough de-
scription of meshless methods, and in particular SPH, can be found in Omidvar (2010). Here,
a SPH method was successfully implemented for the investigation of wave loading on a wave
energy converter.
In a grid-based model the Navier-Stokes equations by themselves are unable to fully de-
scribe the interface of a two-phase flow. Therefore, an additional equation needs to be in-
troduced. Methods for describing the free surface in multiphase flows can be categorized
as Lagrangian-type surface methods and Eulerian-type volume methods, also referred to as
surface tracking and surface capturing methods. A comprehensive analysis of these methods
is given by Gopala and van Wachem (2008).
The surface tracking methods are based on a representation of the free surface by spe-
cial marker points, this allows for explicit tracking of the interface. Interpolation is used to
approximate the points between marker points. The advantage of this approach is that the
position of the interface is known and that the interface remains sharp as it is advected across
the domain. However, these models experience problems when simulating coalescence and
breakup of the interface, so, for instance when a droplet of water is dropped into a bucket
filled with water, or more apt, when wave spray is caused by a large wave colliding with an
offshore structure.The most widely used surface tracking method is the level-set method de-
veloped by Osher and Sethian (1988). Here, the interface is defined as a zero level of a signed
distance function from the interface. To distinguish between the fluids on either side of the
interface a negative sign is attached to the distance function for one of the fluids. The level-set
methods are conceptually simple, relatively easy to implement and yielding accurate results.
However, the method is known to experience loss of mass when the interface is significantly
deformed or when there is considerable vorticity.
In contrast to surface tracking methods, the surface capturing methods support an im-
plicit tracking of the free surface of a two-phase flow. These approaches mark the fluid on
either side of the interface, either by massless particles or by an indicator function. One of
the drawbacks of these techniques is that the exact position of the interface is not known ex-
plicitly and special techniques are needed to reconstruct the interface. The most widely used
surface capturing method is the Volume of Fluid (VoF) method from Hirt and Nichols (1981).
Here, a scalar indicator function between zero and one, known as volume fraction is used to
distinguish between two different fluids. The free surface is treated as the transitional layer
between the two fluids. Due to the diffusive nature of this method, special interface compres-
sion techniques are developed to maintain the sharpness of the interface. A major advantage
of this method is that, in contrary to the level-set method, it is mass conserving and a far
1.3. N UMERICAL SIMULATION OF WAVE - STRUCTURE INTERACTION 13

more robust approach to when it comes to the simulation of coalescence and breakup of the
interface.

1.3.3. N AVIER -S TOKES /VOF SOLVERS


Navier-Stokes-type models coupled with a VOF method for capturing the free surface are a
commonly seen combination in the field of fluid mechanics. They form one of the more
robust methods for simulating the highly nonlinear phenomena caused by fluid-structure
interaction during extreme wave events. A couple of state-of-the-art commercial software
CFD packages are based on the combination of these two models. Note that these software
packages are generally not single purpose, but instead designed for a broader spectra of fluid
mechanics problems.
There are many researches in which commercial software was utilised for the analysis of
floating offshore structures. An excellent example is the analysis of a point absorber wave
energy conversion system, performed by Yu and Li (2013), using STAR-CCM+, a commercial
CFD package, developed by CD-Adapco. Here, a dynamic analysis was performed on the
predominant heave motion of the wave energy converter. ComFLOW is another commercial
state-of-the-art, Navier-Stokes/VOF solver. An extensive review of this model was given by
(Wellens, 2012), where an analysis is performed on an offshore semi-submersible subjected
to waves. The unique feature of ComFLOW is that this method works with a single phase flow.
In addition to commercial software packages, some open-source models are available
which can be of great scientific value if they have a progressive community. Open Field Op-
eration And Manipulation (OpenFOAM®), is an excellent example of such an open source
CFD package. The standard OpenFOAM® software package is distributed with a large num-
ber of solvers and utilities to cover a wide range of problems. Due to the open-source nature
of OpenFOAM®, it is possible for users to write their own codes and solvers for their specific
problems or to modify the existing solvers. Recently the waves2Foam toolbox was developed
by Jacobsen et al. (2012), here, the standard two-phase incompressible Navier-Stokes/VOF
solver was extended with a method to generate and absorb waves. A relaxation zone tech-
nique known as active sponge layer has been applied to the library as well as a large range
of different wave theories. The toolbox is readily available as waveFoam, an extension of
the OpenFOAM® basic interFoam solver. The capability to generate, propagate and absorb
waves has been shown in Jacobsen et al. (2012). This toolbox was also used in an exten-
sive analysis of wave forcing from steep and fully nonlinear waves on a bottom-mounted
offshore wind turbines by Paulsen et al. (2014a). It should be noted that, although the accu-
racy with which the model can propagate waves, determine wave loading and simulate highly
nonlinear wave-structure interaction was actively researched over the last decade, the wave-
structure interaction with respect to floating structures has not been thoroughly examined.

1.3.4. E FFICIENT DOMAIN DECOMPOSITION


Although these Navier-Stokes/VOF models are proven to be very accurate in simulating waves
and wave-structure interaction, the engineering problems which can be investigated are lim-
ited in space and time because in terms of demand of computational power, CFD is an expen-
sive engineering tool. A method with which the computational time can be reduced, without
losing the accuracy in the wave-structure interaction, is to apply a domain decomposition
strategy. Two types of domain decomposition strategies exist: two-way coupling, where in-
formation propagates both from the outer to the inner domain and vice versa, or, one-way
14 1. I NTRODUCTION

coupling, where information only propagates from the outer to the inner domain.The ad-
vantage of one-way coupling is its simplicity and the efficiency with which the coupling be-
tween fully nonlinear models can be established. Recently a fully validated one-way domain
decomposed model was presented by Paulsen et al. (2014b). Here, the fully nonlinear three-
dimensional potential flow model, OceanWave3D, from Engsig-Karup et al. (2009), was cou-
pled with the fully nonlinear Navier-Stokes/VOF solver through the relaxation zones provided
by waves2Foam. The potential flow model can be used to efficiently simulate realistic sea
states while the computationally more expensive Navier-Stokes solvers simulates the highly
nonlinear flow effects involved in the wave-structure interaction.

1.4. T HIS T HESIS


As follows from the discussion in Section 1.1, offshore floating wind can be a promising so-
lution to harvest the wind resources further out at sea. However, in order make it possible to
develop full scale FOWT farms in the near future, reliable and purpose oriented design meth-
ods are required to already reduce uncertainties and risks in the design phase. There are
still many uncertainties with respect to the behaviour of FOWT and their mooring systems
in extreme wave events. These types of structures will be deployed in intermediate water
depths, where the waves during extreme events will be steeper than at deep water locations.
The conventional Boussinesq- and Morison-type engineering tools are insufficient to accu-
rately model the interaction of these extreme waves with a moored floating structure in order
to investigate strongly nonlinear phenomena such as: ULS design loads from (near) break-
ing waves, local impact loads on sensitive structural members, peak forces in mooring and
anchor systems, damping characteristics and diffraction including nonlinear free surface ef-
fects. Highly expensive physical model test campaigns are at present the only valid method
to investigate these effects.
In order to reduce design uncertainties and to speed-up design iterations, an accurate
and well validated numerical model is essential. For such a model to be considered a valu-
able supplement or alternative to physical model testing it shall as a minimum be capable
of describing: strongly nonlinear breaking waves, local impact pressures and wave run-up,
rigid body motion of a moored floating body and loads in the mooring system during wave
impacts. Furthermore, in order to ensure a sufficiently high impact of such a model on the
future development of floating offshore structures, the model should be: geometrical flexible
in order to increase pertinency, fully parallelized for multi-core systems in order to enhance
computational effectiveness, open-source in order to increase accessibility and last but not
least it should be easy to couple with other flow-, soil-, and structural solvers in order to stim-
ulate an integrated design method. The main objective and aim of this thesis is:

"To establish a well validated fully nonlinear numerical wave tank for the simulation
of complex fluid-structure interaction of moored floating offshore structures."

There are a limited number of frameworks available to establish a numerical wave tank that
meets the established requirements. The most suitable models, as discussed in Section 1.3,
are generally based on the Navier-Stokes equation coupled with a geometrical flexible scheme
for describing the free surface. One model, which fits the proposed requirements and stands
out for this research is OpenFOAM®. This widely used open-source CFD library is capable
of solving free surface Newtonian flows by using the Navier-Stokes equations coupled with a
1.5. S TRUCTURE OF THIS THESIS 15

volume of fluid method for tracking the free surface.


In this thesis, the fully nonlinear numerical fluid-structure interaction solvers, interFoam
and interDyMFoam, from the open-source CFD library OpenFOAM® were validated. Please
note that both of these solvers were extended with waves2Foam, a wave generation and
absorption toolbox developed by Jacobsen et al. (2012). Although technological advance-
ments have greatly increased the computational power, the effort involved in running such
fully nonlinear CFD simulations is still substantial and the computational time remains high.
Therefore, the present research explores an efficient domain decomposed strategy to keep
the computational efforts to an acceptable level. Within the scope of this research special
attention will be paid to the following main research questions:

1. Can the Navier-Stokes/VOF solver provide an accurate and reliable solution for nonlin-
ear flow problems?
2. Is the Navier-Stokes/VOF solver capable of accurately computing:
• the propagation of a fully nonlinear wave in intermediate water depth?
• the loads on a structure induced by incoming waves?
• the motions of a floating structure?
• the mooring system of a floating structure?

3. Is it possible to effectively reduce the computational effort by:


• applying grid refinement in the computational domain?
• implementing a domain decomposition strategy?

1.5. S TRUCTURE OF THIS THESIS


The structure of this document is as follows. A review of the utilized numerical models will
be provided in Chapter 2. This chapter includes the description of the general algorithms,
governing flow equations, boundary conditions, wave generation and absorption zones, dy-
namic mesh capabilities and mooring line implementations. Four different two-dimensional
validation cases will be elaborated on in Chapter 3, this is followed by a short discussion, in
Chapter 4, on the effects of the local mesh refinement and automated mesh generation tools,
available in the standard distribution of OpenFOAM®. In Chapter 5 the model is used to sim-
ulate three-dimensional free and moored decay simulations of a FOWT, which are compared
to physical experiments performed in the MARIN basin. The capabilities of the numerical
model will be put to the test in a final proof-of-concept simulation, in Chapter 6, where a
moored FOWT is subjected to irregular. This thesis concludes with a discussion of the pre-
sented work in Chapter 7.
2
T HE NUMERICAL MODELS

2.1. I NTRODUCTION
This section gives a description of the numerical models that were utilised in this research.
To begin with, the fully nonlinear Navier-Stokes/VOF solver used for the accurate simula-
tion of the complex free surface flows and fluid-structure interactions is presented in Sec-
tion 2.2. Furthermore, the combination of this Navier-Stokes/VOF solver with a dynamic
motion solver is treated in Section 2.3. This section also discusses the implementation of
restraints to the floating structure in the form of catenary mooring lines and linear springs.
On top of that, Section 2.4 gives a brief introduction to the potential flow solver which can
be used for fast and accurate reproduction the wave fields. The numerical models applied
in this research are open-source and thus in principle freely available as part of the Open-
FOAM® and waves2Foam framework. It should be noted that the implementation of the
waves2Foam toolbox in the coupled Navier-Stokes/VOF and dynamic motion solver is not
readily available, however, the implementation is relatively trivial and documentation on this
process is available. Furthermore, the implementation of the catenary mooring lines, as pre-
sented in Equation 2.3.5, is developed by Niels Jacobsen (Researcher/advisor at Deltares) and
presently not publicly available.

2.2. I NTER F OAM / WAVE F OAM SOLVER


This section elaborates on the waveFoam solver used to accurately describe the fluid-struc-
ture interaction for situations where no movement of the structure is required. The wave-
Foam solver, is based on the interFoam solver provided by the open-source CFD-toolbox
OpenFOAM®, version 2.3.1., which was extended with the implementation of the wave gen-
eration and absorption toolbox, waves2Foam, developed by Jacobsen et al. (2012). The wave-
Foam solver utilises the two-phase incompressible Navier-Stokes equations in combination
with a VOF-surface capturing scheme (Hirt and Nichols, 1981) to compute fluid-structure
interaction. The focus of this section will be on the governing equations, the boundary con-
ditions and the implementation of wave generation and absorption zones.

17
18 2. T HE NUMERICAL MODELS

2.2.1. G OVERNING EQUATIONS N AVIER -S TOKES /VOF


The governing equations, used in the Navier-Stokes/VOF solver waveFoam, for conservation
of mass and momentum of an incompressible flow of air and water are given by

∇ · u = 0, (2.1)
∂ρu
+ ρ∇(∇ · u)u = −∇p ∗ + (g · x)∇ρ + ρ∇ · (µ∇u), (2.2)
∂t
where ∇ = (∂x , ∂ y , ∂z ) is the three-dimensional gradient operator, u = (u, v, w) is the velocity
field in Cartesian coordinates, g is the gravitational acceleration and p ∗ is pressure in excess
of the hydrostatic pressure, which relates to the total pressure, p, by

p ∗ = p − ρ(g · x). (2.3)

Furthermore, the local density, ρ, and viscosity, µ, are given by the water volume fraction, α,
consistent with

ρ = αρ w at er + ρ ai r (1 − α), (2.4)
µ = αµw at er + µai r (1 − α), (2.5)

where α is zero for air, one for water and a mixture of the two for all intermediate values.
OpenFOAM® uses a VOF method for tracking the air-water interface. After obtaining the
velocity field by solving equations (2.1) and (2.2) for the two-phase flow of air and water, the
VOF method (Hirt and Nichols, 1981) can be used to advance the α field in time with the
following scalar advection equation

∂α
+ ∇ · uα = 0. (2.6)
∂t

Using a standard finite-volume approximation for solving the hyperbolic advection equa-
tion (2.6) would lead to significant smearing of the interface. To reduce this, OpenCFD®
extended the VOF method with an interface compression term. This is thoroughly discussed
in Berberović et al. (2009). The interface compression term, which is the last term on the
left-hand side of the following equation

∂α
+ ∇ · uα + ∇ · ur α(1 − α) = 0, (2.7)
∂t

effectively reduces smearing of the air-water interface. The term is only active in the vicin-
ity of the interface, 0 < α < 1, where its strength is governed by the relative velocity, u r . It is
stressed that even though u r has the dimension of m/s, it lacks any physical meaning. To en-
sure stability a multi-dimensional flux limited scheme (MULES) is used for solving the scalar
advection equation (2.7).
In spite of the addition of the interface compression term to the volume of fluid method,
the air-water interface can be smeared over several computational cells. Consequently these
cells have a water volume fraction of 0 < α < 1, which means that there is no clear location of
the interface. If a free surface location is required, this can be defined by

L ai r
w at er (α) = {x|α(x) = 0.5}, (2.8)
2.2. I NTER F OAM / WAVE F OAM SOLVER 19

Figure 2.1: Schematic representation of the computational domain in the Navier-Stokes/VOF solver, where I and
II are relaxation zones implemented with the waves2Foam toobox and all boundary surfaces were denominated
with names used in the present research.

which defines the free surface by the iso-contour where α = 0.5. Alternatively, the loca-
tion of the free surface can be determined by making use of the wave gauge functionality
in waves2Foam. One or more wave gauges can be placed at certain positions in the numeri-
cal domain, these determine the location of the free surface, η, relative to the still water level,
according to
Z x1
η= αdz − h. (2.9)
x0
Here, x0 and x1 are the user defined start and end points of a vertical line over which the α
field is integrated and h is the water depth at the still water level condition. This approach
yields a consistent value of the location of the free surface and is used throughout the present
work for determination of the interface at specific location. The former method, using the
iso-contours where α = 0.5, is used purely for visualisation purposes.

2.2.2. B OUNDARY CONDITIONS


In order to solve the governing equations of the Navier-Stokes/VOF solver, boundary con-
ditions need to be specified for all surfaces in the numerical domain. Before going into the
allocation of particular boundary conditions, the general denomination of the boundary sur-
faces is given in a schematic representation of the numerical domain, see Figure 2.1. It should
be noted that a structure will be present inside the numerical domain for a greater part of the
numerical computations, the boundary surface of the structure is generally denominated as
the floating object.
At the inlet and outlet boundary surfaces the velocities, u, and the α field are given by a
closed form analytic wave theory or by the potential flow solver (see Section 2.4). Assuming
the wave impacts are inertia dominated, a slip condition is applied to the surfaces of, fixed
or moving, structures (e.g. horizontal cylinders in Chapter 3 and 4, or the floating offshore
wind turbine structure in Chapter 5 and 6. The viscous boundary layer at the seabed is also
neglected, since typically the propagation of waves with a CFD-solver is limited to only a cou-
ple of wavelengths and in such simulations the influence of the seabed on the wave forcing is
presumed small. The viscous boundary layer of side-walls of the three-dimensional numer-
ical wave-tank, as presented in Chapter 5 and 6, are also subjected to a slip condition, while
for the two-dimensional computations, as presented in Chapter 3 and 4, an empty condition
was applied to these walls. It may be noted that, by imposing a slip condition, the viscous ef-
fects are artificially suppressed and although OpenFOAM® supports turbulence modelling,
20 2. T HE NUMERICAL MODELS

Figure 2.2: Schematic representation of the spatial weighting factor, χ(ξ), applied to match the target and the
computed values in the relaxation zones. Here, I and II can be interpreted as an inlet and outlet respectively.

it is deliberately excluded in the present work. At the atmosphere boundary layer of the nu-
merical domain, an atmospheric boundary condition is applied for the velocity, u, and the α
field. This means that air and water are allowed to leave the numerical domain, while only
air is allowed to flow back in. Furthermore, the total pressure at the atmosphere boundary is
equal to zero.

2.2.3. R ELAXATION ZONES


In the wave generation toolbox, waves2Foam, developed by Jacobsen et al. (2012), relaxation
zones were developed based on an extension of the relaxation technique by Mayer et al.
(1998). These relaxation zones were implemented to prevent wave reflection from the outlet
boundary and also to prevent internally reflected waves, e.g. by the presence of a structure,
to interfere with the wave generation at the inlet boundary. An arbitrary number of relaxation
zones can be defined, which can be characterised as either an inlet or outlet. In the present
work all implemented relaxation zones were rectangular shaped, nevertheless circular ring
shaped relaxation zones can also be implemented and with some minor changes to the code
of waves2Foam it would also be possible to define other geometrical shapes. In the relaxation
zones the velocity and α field are updated every time step according to

ψ = χψt ar g et + (1 − χ)ψcomput ed , ψ ∈ {uH , w, α}, (2.10)

where the target solution, ψt ar g et , can be specified by one of the wave theories available in
waves2Foam or by the potential flow solver, presented in Section 2.4. The computed solution,
ψcomput ed , on the other hand are the numerically computed velocity and α values, obtained
from the Navier-Stokes/VOF. The weight of the each of the two solutions is determined by

exp(ξβ ) − 1
χ(ξ) = 1 − , (2.11)
exp(1) − 1
where χ is the local coordinate in the relaxation zone, which, as Figure 2.2 indicates, is
one at the outer end and zero on the inner end of the relaxation zone. The shape factor, β,
can be chosen arbitrarily, however in the present work the default value, β = 3.5, was used.

2.2.4. A LGORITHM WAVE F OAM


The following algorithm gives a schematic overview of the process sequence of the waveFoam
solver. The actual code is presented in Appendix A. It may be noted that the only difference
between the interFoam solver and the waveFoam solver is the highlighted relaxation of the
alpha and velocity field.
2.3. I NTER DY MF OAM / WAVE DY MF OAM SOLVER 21

. Load case data


. Initialise waveDyMFoam solvers
while runTime
. load run controls
. set initial time step
while PIMPLE loop
if {first pimple loop} alpha
. solve VOF equation
. relaxation alpha & velocity field
. update ρ and µ
. update velocity field
while pressure loop
. update pressure field
end pressure loop
end PIMPLE loop
end runTime

2.3. I NTER DY MF OAM / WAVE DY MF OAM SOLVER


This section shortly discusses features of the waveDyMFoam solver, which is utilised in the
present work to describe the fluid-structure interaction for cases where motion of the struc-
ture is required. The waveDyMFoam solver, is based on the interDyMFoam solver provided
by the open-source CFD-toolbox OpenFOAM®, version 2.3.1. This solver couples the Navier-
Stokes/VOF solver with a dynamic motion solver. For the present research, the OpenFOAM®
standard interDyMFoam solver was extended with the wave generation and absorption tool-
box, waves2Foam, developed by Jacobsen et al. (2012), and the restraints for floating struc-
tures developed by Niels Jacobsen (Researcher/advisor at Deltares). This section will focus
on the dynamic motion solver, the dynamic mesh handling and the implementation of the
motion restraints.

2.3.1. M OTIONS OF A BODY


In order to simulate the flow-dependent motion response of a floating structure, for instance
in the heave decay simulation of the two-dimensional horizontal cylinder discussed in Sec-
tion 3.5, the Navier-Stokes/VOF solver has to be coupled with a six degrees of freedom (6-
DOF) motion solver.
The term six degrees of freedom refers to the freedom of motion of a body. If a body has
six degrees of freedom, it is free to translate along and rotate around the x, y and z axes, see
Figure 2.3. It is also possible to restrict the number of degrees of freedom to, for instance,
just translation along the vertical direction as is done to single out heave displacement in the
two-dimensional heave decay simulation discussed in Section 3.5.
The present only work considers rigid body motion, of which the accelerations, veloci-
ties and displacements can be calculated from the forces and moments on the body. The
waveDyMFoam solver extracts the total force on and moment around the center of gravity
(COG) by integrating all the separate forces and moments over the body according to
22 2. T HE NUMERICAL MODELS

Figure 2.3: Axis describing six degrees of freedom, three translations along and three rotations around the x, y and
z axes.

Z
F= F f l ow + Fext , (2.12)
B
Z
M= M f l ow + Mext . (2.13)
B
Here, F f l ow and M f l ow are determined from integrating the pressures obtained by solving
the Navier-Stokes equation (2.2) and the continuity equation (2.1). The Fext and Mext are
considered to be external forces and moments, an example of an external force would be the
gravitational force, which acts on the COG of the body. Other external forces and moments
can be restraining forces due to the anchoring of a floating structure with the use of mooring
lines, a discussion on the implementation of these follows in Subsection 2.3.5.

2.3.2. U NDER - RELAXATION


As introduced above, the hydrodynamics and the motions of the structure can be decoupled
in the numerical simulation of floating structures. The concept of negative added mass in
weakly coupled numerical methods can give rise to instabilities, this phenomenon is thor-
oughly discussed in the work from Seng (2012) and Dunbar et al. (2015). Here, strong cou-
pling between the Navier-Stokes/VOF solver and the 6-DOF solver was established, which
successfully improved the stability of the numerical model. It should be noted that their find-
ings were based on an older version of OpenFOAM®.
In the newer version of OpenFOAM® that is used in the present research an under-relax-
ation method is available to counteract the stability issues. Under-relaxation methods make
use of an under-relaxation factor, a comprehensive evaluation of the influence of this under-
relaxation factor is presented in Ferziger and Peric (2012). The under-relaxation method is
embedded in the 6-DOF motion solver, here, the acceleration of the COG of the structure,
which is calculated from the forces, is ’relaxed’ by an ’accelerationRelaxation’ factor, f , ac-
cording to

a 2∗ = f a 2 + (1 − f )a 1 , (2.14)
a 3∗ = f a 3 + (1 − f )a 2∗ , (2.15)
a 3∗ = f a 3 + (1 − f )( f a 2 + (1 − f )a 1 ), (2.16)
a 3∗ = f a 3 + f a 2 − f 2 a 2 + (1 − f )2 a 1 , (2.17)
2.3. I NTER DY MF OAM / WAVE DY MF OAM SOLVER 23

where, a n∗ , is the under-relaxed acceleration of the COG of the structure at the instantaneous
time step. This is defined by the acceleration calculated from the forces and moments mul-
tiplied by the ’accelerationRelaxation’ factor plus the under-relaxed acceleration of the pre-

vious time step, a n−1 , multiplied with one minus the ’accelerationRelaxation’ factor. Note
that the relaxation of a, only starts from the second time step, this is defined in the full code
of the 6-DOF rigid body motion solver and the embedding of the under-relaxation method
which can be found in Appendix C. This implies that if, f = 0.0, the under-relaxation pre-
vents the structure from moving at all, since a 2∗ = f a 2 + (1 − f )a 1 = 0 because a 1 = 0. While
the implementation of the under-relaxation method turns the acceleration into a total vari-
ation diminishing (TVD) property, it should be noted that by doing so it also introduces a
diffusive term to the numerical model which can have a negative effect on the convergence
rate of the solution. Throughout the present research an ’accelerationRelaxation’ factor of
0.5 is used, this value was adopted from default settings in the ’floatingBlock’ tutorial from
OpenFOAM®.

2.3.3. DYNAMIC MESH HANDLING


After the displacement of the body is determined, the mesh has to be adapted in order to
comply with this motion. Two fundamentally different methods can be distinguished for
handling moving boundary surfaces. One features a dynamic mesh that allows for topologi-
cal changes, which means that the number of points, faces, and cells in the mesh are allowed
to change in order to concur with the motion of the body. The other method, is dynamic mesh
deformation, which, as the name suggests, allows the mesh to deform while the number of
points, faces and cells remains constant. The latter of the two is applied in the waveDyM-
Foam solver since the former involves data mapping between time-steps which can be asso-
ciated with numerical errors.
Different mesh deformation algorithms exist that can be used to compute the motion of
the internal mesh points based on the movement of the boundary surface. The work of Bos
et al. (2013) gives a thorough discussion on the implementation of a state-of-the-art defor-
mation algorithm, the radial basis function interpolation, in OpenFOAM®. The mesh de-
formation algorithm in the version 2.3.1. of OpenFOAM® is based on the mathematically
related spherical linear interpolation (Slerp) function, as presented in Subsection 2.3.4. This
technique is thoroughly discussed in the work of Maruyama et al. (2012).
When a dynamic mesh problem is considered, the shape of the computational domain is
varying in time, a schematic representation of this is presented in Figure 2.4. Here, the nu-
merical domain Ω, alters its shape, during a time interval ∆t , to form a new configuration Ω0 .
In this transformation, a distinction can be made between the motion of the boundary points
and the internal mesh points. The displacement of the boundary surface, B , is determined by
the forces and moments on that boundary, while the internal mesh points need to be moved
as a reaction to the motion of the boundary points. Here, the mesh deformation algorithm
has the responsibility to maintain the quality and validity of the mesh.
In order to enable the movement of the boundary surface, B , and preserve the mesh qual-
ity, two different areas are defined around the moving boundary surface. The outer radius de-
fines an area in which numerical cells are allowed to deform according to the described mo-
tion of the body. A smaller area is defined by the inner radius, in the area between the bound-
ary surface and the inner radius, the computational cells translate and rotate according to the
motion of the boundary surface without deforming, thus preserving the initial mesh orthog-
24 2. T HE NUMERICAL MODELS

Figure 2.4: Mesh deformation in the computational domain, Ω, during the motion of a body, B . Here, (a) is the
initial configuration and (b) is the deformed configuration.

onality. It may be noted that the internal point motion influences the solution only through
discretisation errors (Ferziger and Peric, 2012).

2.3.4. R OTATIONS AND TRANSLATIONS


Older versions of OpenFOAM® used generic motion solvers based on solving an elliptic equa-
tion for the displacement. These motion solvers suffered from problems in terms of mesh
quality, causing numerical instability. One of the most important entities for the mesh qual-
ity is the orthogonality of the grid lines, which is important for accurate flux calculations.
Version 2.3.1. of OpenFOAM® uses a specialised 6-DOF motion solver that handles the
the motion of the body and the mesh deformation by interpolating the translations and rota-
tions using Slerp. These translations and rotations are expressed in quaternions en septernions,
where a septerion consists of a vector and a quaternion. This subsection will give an insight
in the use of Slerp and quaternions, more thorough studies on the subject can be found in the
works by Shoemake (1985), Dam et al. (1998), Samareh (2002) and Maruyama et al. (2012).

R OTATION USING QUATERNIONS


Three different types of methods can be distinguished for the representations of rotation:
Euler angels, rotation matrices and quaternions. The quaternion formulation is used in the
waveDyMFoam solver, quaternions can be seen as a three-dimensional extension of complex
numbers, also referred to as hyper complex numbers. They were first described by Hamilton
(1844). A quaternion, Q, can be represented by a vector in a four-dimensional hypersphere
according to

Q(w, v) = q 1 + i q 2 + j q 3 + kq 4 , w = q1 , v = (q 2 , q 3 , q 4 ), (2.18)

where i , j and k are imaginary units, w is the real scalar and v the imaginary vector, see
Figure 2.5a for a schematic interpretation. The magnitude of the quaternion on the unit hy-
persphere is defined by
q
|Q| = q 12 + q 22 + q 32 + q 42 = 1, (2.19)

and the multiplication of two quaternions is defined by

Q12 (w 12 , v12 ) = [w 1 w 2 − v1 · v2 , w 1 v2 + w 2 v1 + v1 × v2 ],

where it should be noted that the order of multiplication matters for the result, i.e.
2.3. I NTER DY MF OAM / WAVE DY MF OAM SOLVER 25

Figure 2.5: Schematic interpretations of the fourth dimension, where (a) is the four dimensional space in which a
quaternion can be defined and (b) is the interpretation of rotation of a point by a quaternion.

Q12 = Q1 Q2 6= Q2 Q1 . (2.20)

A quaternion can also be represented as a rotation of a point by a certain angle, θ, about a


unit vector, n̂, according to the following equation

θ θ θ θ
· µ ¶ µ ¶¸ µ ¶ µ ¶
Q(w, v) = cos , n̂ sin , w = cos , v = n̂ sin , (2.21)
2 2 2 2
where the inverse of the quaternion is defined by

θ θ
· µ ¶ µ ¶¸
−1
Q (w, −v) = cos , −n̂ sin . (2.22)
2 2
Using equations (2.21) and (2.22), a point p = (x, y, z), expressed as quaternion P(0, p) can be
rotated over a quaternion Q according to

Pr ot at ed = QPQ−1 = P0 Q−1 , (2.23)

where it may be noted that the first operation in equation (2.23), QP = P0 , can be seen as the
first half of the rotation, where one enters the fourth dimension, and the second operation,
P0 Q−1 , the other half of the rotation, which cancels out the rotation in the fourth dimension,
this is presented in a schematic manner in Figure 2.5(b).

I NTERPOLATION
Slerp is a powerful interpolation method used for the interpolation of quaternions. They
trace the minimum distance path between quaternions on the hypersphere. Here, a brief
introduction to linear interpolation (Lerp) and spherical linear interpolation (Slerp) are given,
where Lerp is used to interpolate the translation and Slerp to interpolate the rotation in body
motion and mesh deformation.

L INEAR I NTERPOLATION (L ERP )


Lerp is the most widely applied technique for interpolation of values. The simplest example
of Lerp, as presented in Figure 2.6, is interpolation using two weighted values, 1 − α and α,
according to the following formula
26 2. T HE NUMERICAL MODELS

Figure 2.6: Schematic interpretations of the different interpolation methods, where (a) is linear interpolation in a
three-dimensional space and (b) is spherical linear interpolation in a four-dimensional space.

tl er p (t1 , t2 ; α) = (1 − α)t1 + αt2 . (2.24)

This type of interpolation is used for the interpolation of the translation vectors. Further-
more, the process of interpolation is commutative, which means it can be generalised for any
number of interpolations according to
,
n t n 1
X i X
tl er p (t1 , t2 , · · · , n; l 1 , l 2 , · · · , l n ) = . (2.25)
l
i =1 i i =1 l i

S PHERICAL LINEAR INTERPOLATION (S LERP )


Slerp is an interpolation method that divides the angle between two vectors by an arbitrary
ratio. Slerp was introduced as a method for interpolation of quaternions by Shoemake (1985),
it defines the minimum distance trace between two points on the hypersphere. Two quater-
nions can be interpolated with Slerp, see Figure 2.6, using two weighted values, 1 − α and α,
according to the following formula

sin [(1 − α)Θ] sin [αΘ]


Qsl er p (Q1 , Q2 ; α) = Q1 + Q2 , (2.26)
sin Θ sin Θ
where Θ is defined as the angle between two quaternions, θ = cos (Q1 · Q2 ). It may be noted
that as Θ → 0, this formula reduces to the formula for linear interpolation, equation (2.25).

2.3.5. R ESTRAINTS FOR MOVING BOUNDARY SURFACES


This subsection discusses the implementation of restraints for moving boundary surfaces
provided by Niels Jacobsen (Researcher/advisor at Deltares). It should be noted that no offi-
cial documentation of this implementation was available at present.
Two types of restraints have been implemented on version 2.3.1 of OpenFOAM®, a linear
spring and a simple catenary type mooring line. It should be noted that this simple mooring
line implementation does not take the interaction with the fluid into account, thus it is not
influenced by waves or current. The restraints can be used to connect a floating body to a
fixed anchor point, which, for the mooring line, is assumed to be located on a horizontal sea
bed. Another application, although not investigated in this paper, would be to interconnect
two floating bodies, for instance to represent a towing line connection between a tug boat
2.3. I NTER DY MF OAM / WAVE DY MF OAM SOLVER 27

Figure 2.7: Schematic representation a the catenary mooring line, where p0 is the anchor point, p1 is the attach-
ment point to the floating object and pt is the touchdown point of the line.

and a floating cargo. Each of these restraints gives rise to forces in the attachment points.
This subsection discusses the derivation of these forces, as well as a brief description of the
restraints.

L INEAR SPRING
The linear spring is implemented without any damping and the line is given a simple constant
stiffness, k, and a rest length, l r . The spring is defined such that the force in the spring is zero
when the spring is at rest, i.e. has a length equal to l r . The force exerted by the linear spring
can be calculated according to
r
F0 = −k (krk2 − l r ) = −F1 , (2.27)
krk2
where F0 and F1 are the spring forces and r = p0 − p1 is the distance between the two connec-
tion points.

M OORING LINE
The mooring line implementation in the present work is used for anchoring a floating struc-
ture to the seabed, see Figure 2.7, where it is assumed that the seabed is horizontal. Three
different mooring line states can be defined; simple state, resting state and hanging state,
see Figure 2.8. These states and their corresponding forces in the attachment points of the
mooring line are described in the following section.

S IMPLE STATE
In the simple state, see Figure 2.8a, the mooring line is completely in suspension, i.e. the an-
chor point, p0 , is equal to the touchdown point, pt . The implementation of the mooring line
assumes that the line cannot break, while the distance between the connected objects can
exceed the defined length of the line, l c . This means that the state of the line can either be or-
dinary or length exceeded. The ordinary formulation follows the well know resistance force in
a catenary line between two attachment points, where the distance between the attachment
points is less than the length of the line, see e.g. Krenk (2001). The horizontal restraining
force, F H , is a function of the distance between the two attachment points following

s2 − d 2
FH = m sub g , (2.28)
2d
where s is the instantaneous length of the mooring line, which in this case is equal to l c , d
is the vertical distance between the attachment points and m sub is the submerged weight of
the mooring line. The resultant force, F T , can than be found according to
28 2. T HE NUMERICAL MODELS

Figure 2.8: Schematic representation of the three different states of the catenary mooring line, where (a) is the
simple state, (b) the resting state and (c) the hanging state. furthermore, p0 is the anchor point, p1 is the attach-
ment point to the floating object and pt is the touchdown point of the line.

F T = F H + m sub g l (2.29)
where l is the vertical distance between the attachment point p1 and the touchdown point,
pt , which, in this case of the simple state, is equal to p0 .
When the distance between the two attachment points, krk2 , is larger than the 0.999l c ,
the line is treated as a linear spring. The stiffness of the line, k c , is based on the magnitude
of the restraining force in the catenary line at the limit of its length and can be calculated
according to

F0.999 − F0.998
kc = , (2.30)
0.001l c
where F0.999 and F0.998 are the forces in the attachment point p1 , when the line has a length
of respectively 0.999l c and 0.998l c respectively. The force exerted by the mooring line, when
the length of 0.999l c is exceeded, can be calculated by
r
F0 = −k (krk2 − 0.999l r ) + Fcr = −F1 , (2.31)
krk2
where Fcr is the tension in the mooring line at length 0.999l c . It may be noted that the moor-
ing line is fully stretched and that the force Fcr is parallel with r.

R ESTING STATE
The resting state, see Figure 2.8b, is the mooring line state, where the floating structure is so
close to the anchor point that a part of the catenary line will be resting on the sea bed. The
remainder of the cable will behave as a catenary line. In order to evaluate the restoring force
on the floating object in the attachment point, the touchdown point pt is required, so that
the instantaneous length, s, of the suspended part of the mooring line can be determined.
The touchdown point, pt , is evaluated by finding the resting length of the line that ensures
that the slope of the line at the touchdown point is horizontal, i.e.

∂pt
= 0. (2.32)
∂z
Here, the touchdown point, pt , is defined as the resting length away from the anchor point.
This requires iterations in the description of the mooring line. Once the resting length has
been identified, the forces can be found in the attachment point following the catenary the-
ory, see equations (2.28) and (2.29).
2.4. F ULLY NONLINEAR POTENTIAL FLOW SOLVER 29

H ANGING STATE
The hanging state is defined as the state where the floating object is so close to the anchor
point that the piecewise linear configuration of the mooring line is shorter than the length of
the mooring line, see Figure 2.7(c). It is defined that this state takes place when the angle, γ,
between the line at the attachment point, p1 and the horizontal exceeds 88◦ . Only a vertical
force acts in the attachment point, p1 , in this state, which equals the weight of the suspended
part of the line. It may be noted that both in the resting state and the hanging state, the force
in the anchor point, F0 , vanishes.

2.3.6. A LGORITHM WAVE DY MF OAM


The following algorithm gives a schematic overview of the process sequence of the waveDyM-
Foam solver. The actual code is presented in Appendix B. It may be noted that the waveDyM-
Foam solver is basically the waveFoam solver with the addition of highlighted 6-DOF loop.
The only difference with the interDyMFoam solver being the introduction of relaxation zones.

. Load case data


. Initialise waveDyMFoam solvers
while runTime
. load run controls
. set initial time step
while PIMPLE loop
while 6-DOF loop
. extract forces and moments
. move body / deform mesh
. update flux field
end 6-DOF loop
if {first pimple loop} alpha
. solve VOF equation
. relaxation alpha & velocity field
. update ρ and µ
. update velocity field
while pressure loop
. update pressure field
end pressure loop
end PIMPLE loop
end runTime

2.4. F ULLY NONLINEAR POTENTIAL FLOW SOLVER


Recently the Navier-Stokes/VOF solver was coupled with the fully nonlinear potential flow
solver (OceanWave3D), developed by Engsig-Karup et al. (2009), to form a one-way domain
decomposed solver (Paulsen, 2013) (Paulsen et al., 2014b).
A schematic representation of the numerical domain of the domain decomposed solver
is presented in Figure 2.9. In the outer domain, Ω, the fully nonlinear potential flow problem
is solved for the wave motion only. Additionally a smaller inner domain, Γ, is defined to
30 2. T HE NUMERICAL MODELS

Figure 2.9: Schematic representation of the numerical domain, where Ω is the outer domain governed by the
potential flow solver and Γ is the inner domain where the flow is described by the Navier-Stokes/VOF solver.

model wave-structure interaction and structure motions. Here, the Navier-Stokes/VOF solver
is used to compute the velocity and α field. The relaxation zones in the Navier-Stokes/VOF
solver are used to extract information about the wave field from the outer domain, i.e. the
velocity and surface elevation are interpolated from the outer domain, Ω, to the relaxation
zones in the inner domain, Γ. The Navier-Stokes/VOF solver is essentially using the fully
nonlinear potential flow solution as target solution in the relaxation zones instead of one of
the wave theories available in waves2Foam.
Since the potential flow solver is significantly faster and more accurate than the Navier-
Stokes/VOF solver, due to stable higher order discretisation, it becomes possible to generate
larger domains and longer time series. The coupling is set-up in a way that the potential
flow solver can run a time series until a particular moment of interest, e.g. an extreme wave
event, and only then start the computationally more expensive Navier-Stokes/VOF solver,
after which there is runtime coupling between the two, see Subsection 2.4.2. This coupling
supports parallel computations and utilises shared memory to improve performance.

2.4.1. G OVERNING EQUATIONS


Potential flow assumes the fluid to be inviscid and incompressible so that it can be described
by the velocity potential, φ, and the position of the free surface. The physical fluid veloci-
ties can be expressed by (uH , w) = (∇H φ, ∂z φ), where uH = (u, v) and ∇H = (∂x , ∂ y ). The free
surface, η, relative to the still water level, can be evaluated according to the kinematic free
surface condition
∂t η = −∇H η · ∇H φ̃ + w̃(1 + ∇H η · ∇H η), (2.33)
where w̃ = φ(x, η) and φ̃ = φ(x, η) are the velocity and velocity potential at the free surface. By
specifying the pressure at the free surface to be zero, the Bernoulli equation can be rewritten
as the dynamic free surface condition,

∂t φ̃ = −g η − ∇H φ̃ · ∇H φ̃ − w̃ 2 (1 + ∇H η · ∇H η) .
¢
(2.34)
2
The vertical velocity at the free surface, w̃, is required to advance the free surface conditions
in time. For this, the velocity potential in the fluid volume must be determined by solving the
Laplace equation,
∇2H φ + ∂zz φ = 0, −h ≤ z < η, (2.35)
together with the kinematic sea bottom condition,

∂z φ + ∇H h · ∇H φ = 0, z = −h, (2.36)

where h = h(x) is the water depth measured from the seabed to still water level. A detailed de-
scription of the fully nonlinear potential flow solver, one-way domain decomposition strategy
and validation cases can be found in Paulsen (2013) and Paulsen et al. (2014b).
2.4. F ULLY NONLINEAR POTENTIAL FLOW SOLVER 31

2.4.2. A LGORITHM WAVE DY MF OAM / POTENTIAL FLOW


The following algorithm gives a schematic overview of the process sequence of the waveDyM-
Foam solver including a runtime coupling of the potential flow solver.

. Load case data


. Initialise waveDyMFoam solvers
while runTime
. load run controls
. set initial time step
. update wave field
while PIMPLE loop
while 6-DOF loop
. extract forces and moments
. move body / deform mesh
. update flux field
end 6-DOF loop
if {first pimple loop} alpha
. solve VOF equation
. relaxation alpha & velocity field
. update ρ and µ
. update velocity field
while pressure loop
. update pressure field
end pressure loop
end PIMPLE loop
end runTime
3
WAVE PROPAGATION AND
FLUID - STRUCTURE INTERACTION IN 2D

3.1. I NTRODUCTION
This chapter explores the capabilities of the Navier-Stokes/VOF solver in a two-dimensional
domain in terms of wave propagation with and without the presence of an object. For cases
with a static mesh the waveFoam solver is used, whereas the waveDyMFoam solver is used for
the cases where mesh motion is required. For this investigation, several cases of increasing
complexity were modelled and analysed. A grid convergence study on the propagation of a
fully nonlinear stream function wave (Fenton, 1988) is discussed in Section 3.2. Followed by a
comparison between the experimental data from Dixon et al. (1979) and the numerical results
for wave loading on a fixed, partially submerged horizontal cylinder in Section 3.3. In con-
trast to these first cases, the last cases investigate the capabilities of the Navier-Stokes/VOF
solver when used in conjunction with a dynamic mesh. Section 3.4 reports on the generation
of surface waves by the forced heave oscillating of a horizontal cylinder in a finite domain,
where the result is compared with experimental data from Yu and Ursell (1961). The chapter
concludes with the investigation of the heave decay of a free floating horizontal cylinder in
Section 3.5. Here, the numerical result is compared to theoretical work by Maskell and Ursell
(1970) and experimental work by Ito (1977).

3.2. C ONVERGENCE STUDY ON THE PROPAGATION OF A FULLY NONLIN -


EAR REGULAR WAVE

This section evaluates the ability of the Navier-Stokes/VOF solver to accurately describe the
propagation of a fully nonlinear regular stream function wave in intermediate water depth.
This is evaluated in a two-dimensional grid convergence study similar to that of earlier work
presented by Paulsen (2013) and Paulsen et al. (2014a). The aim of such a grid convergence
study is to evaluate whether the solution improves with the anticipated convergence rate
when the spatial discretisation of the domain is changed. The predominant focus of this
section is on the effect of the discretisation of the convective term in the momentum equation
(2.2) by a first order Upwind scheme, also the investigation of a higher order scheme will be
treated.

33
34 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.1: Schematic representation of the numerical domain used for the propagation of a fully non-linear
regular stream function wave, where l = 200m, d = 50m and a = 26m.

3.2.1. N UMERICAL SET- UP


In the evaluation of the convergence rate of the Navier-Stokes/VOF solver, a continuously
propagating fully nonlinear stream function wave is modelled in a two-dimensional domain.
The properties of this nonlinear wave are specified in Table 3.1. The analytical wave profile
was generated according to the stream function wave theory from Fenton (1988). It may be
noted that this analytical solution only applies to a single phase flow of water, whereas the
numerical solver gives the solution for a two-phase flow of water and air. The effect of the air
phase is therefore part of the evaluation.
The convergence study was carried out in a rectangular domain where cyclic boundary
conditions were applied at the inlet and outlet of the domain, as presented in Figure 3.1. By
implementing these cyclic domain boundaries in the direction of propagation, a domain of
only one wavelength could be used and the relaxations zones, used to generate and absorb
waves, become unnecessary. This strategy, adopted from Paulsen et al. (2014a), was imple-
mented to keep the computational time low and eliminate any influence the relaxation zones
might have on the solution. The viscous boundary layer on the seabed was neglected and a
slip condition was applied which is in accordance with the potential flow assumption of the
stream function solution. An open inlet/outlet boundary condition is applied on the atmo-
sphere boundary surface, satisfying that the total pressure is zero. The domain has a length,
l , which is equal to the wavelength, λ. The height of the numerical domain was defined as
the water depth, d , at still water, plus a vertical dimension, a, to accommodate for the free
surface elevation and the air phase. Here, the length of a is equal to one and a half times the
wave height, H . The domain was discretised by n × m cells with a unit aspect ratio, where the
cell size is defined by the number of points per wavelength (p.p.w.) as,
Table 3.1: Wave properties of the fully nonlinear regular wave used for the convergence studies.
Property Symbol Value
Water depth kh 1.571
Wave height kH 0.534
p
Wave period T/ gk 20.37
p
Eulerian current velocity US / g k 0.00
Number of Fourier modes N 32
Wave number k 3.14 · 10−2 m−1
3.2. C ONVERGENCE STUDY ON THE PROPAGATION OF A FULLY NONLINEAR REGULAR WAVE 35

Figure 3.2: Visualisation of the numerical simulation of the propagation of a fully nonlinear stream function
wave. Here, the iso-contour with α = 0.5 is used to visualise the free surface at time interval (a) t /T = 0.0 and (b)
t /T = 15. The spatial discretisation used for this simulation was 50 p.p.w.

H
∆= . (3.1)
p.p.w.
The spatial discretisation of the domain is given in Table 3.2 by the number of p.p.w., the
actual cell size in meters and the total number of computational cells in the domain.
The Courant-Friedrichs-Lewy (CFL) condition is utilised, for all the numerical computa-
tions performed throughout this research, to determine the time step. For a two-dimensional
case, the CFL condition is described as

|u x |∆t |u y |∆t
Co = + ≤ C o max , (3.2)
∆x ∆y
where the dimensionless number, C o, is the Courant number, u x,y is the magnitude of the
velocity, ∆t is the time step and ∆x and ∆y are the cell sizes in the x and y direction respec-
tively. To ensure numerical stability the maximum Courant number for an explicit time solver
should have a value between one and zero. This condition ensures that a particle cannot
move further than the length of a computational cell. Unless otherwise specified a Courant
number of C o = 0.20 is used throughout the two-dimensional simulations in Chapter 3 and
4.
A run-time analysis of the numerically computed free surface elevation was done with
the use of 39 uniformly distributed wave gauges, i.e. a wave gauge was placed every 5 meters.
Subsequently the phase error, δθ, and the diffusive error, δη, are determined by comparing
the numerically computed solution with the analytical stream function solution. The diffu-
sive error, i.e. the change in wave height, was calculated using
q
0
δη = 〈(η xi )2 〉 (3.3)
where the mean, 〈·〉, of the difference between the numerically computed free surface eleva-
0
tion and the analytical stream function solution, η xi , is calculated. Since the solution may
also have a phase lag, the analytical stream function solution is shifted over a phase range,
− π2 ≤ θ ≥ π2 , for each time step. The lowest diffusive error indicates the spatial position of the
numerically computed wave and with it the phase error, δθ, was established.

3.2.2. A NALYSIS CONVERGENCE STUDY


The numerical simulation of the propagation of the fully nonlinear stream function wave in
a two-dimensional domain with cyclic boundaries is visualised in Figure 3.2. Here, the free
surface is visualised using the iso-contour with α = 0.5. It can be observed that the wave not
36 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.3: Diffusive error, normalized by the wave height, as a function of time, normalized by the wave period.
The convective term of the momentum equation is discretised with a first order Upwind scheme. The lower two
axes indicate the logarithmic rate of convergence, where (a) represents the convergence rate for 50, 100 and 150
p.p.w. and (b) represents the convergence rate for 100, 150 and 200 p.p.w.

only becomes significantly smaller over time, expressed in this section as the diffusive error,
but a phase shift is also introduced. The diffusive error is depicted as a function of time in
Figure 3.3. Here, it can be observed that the initialisation of the wave field is the same for the
simulations with a spatial resolution of 100, 150 and 200 p.p.w., while the diffusive error is
slightly larger for the simulation where a spatial discretisation of 50 p.p.w. was applied. This
may indicate that the 50 p.p.w. mesh is too coarse to fully grasp the detail of the initialised
wave. In all four simulations it can be seen that the diffusive error significantly increases
over the first half of the first wave period. When the α values are used to visualise the wave
field it can be observed that the fluid interface, between the water and the air phase where
0 < α < 1, is sharply defined at time step zero. These visualisation also showed that, although
the applied VOF method for tracking the free surface includes an interface compression term,
equation (2.7), the interface is slightly smeared during the first half of the first wave period, as
presented in Figure 3.4. This smearing effect is believed to be the main reason for the initial
increase in the diffusive error during the first number of periods. After the initial increase of
the diffusive error, over the wave first period, a more steady small increase is observed for the
next fourteen wave periods. With increasing spatial resolution a reduction in the diffusive
error can be seen and for the finest mesh resolution, with 200 p.p.w., the error remains in
the order of O(10−2 ) for the first 15 wave periods. It may be noted that the propagation of
waves over 15 wavelengths with CFD software like OpenFOAM® is without a doubt beyond
Table 3.2: Spatial discretisation and total number of computational cells of the numerical domain used for the
grid convergence cases.
Points per wavelength Cell size, ∆ [m] Computational cells
50 4.00 950
100 2.00 3800
150 1.33 8550
200 1.00 15200
3.2. C ONVERGENCE STUDY ON THE PROPAGATION OF A FULLY NONLINEAR REGULAR WAVE 37

Figure 3.4: Visualisation of the iso-contours with an alpha value equal to 0 < α < 1. Smearing of the interface is
visualised at (a) t /T = 0.0 and (b) t /T = 0.5, where the spatial discretisation of the simulation was 50 p.p.w.

the intended purpose of such software and that for most practical engineering purposes the
propagation will be limited to a couple of wavelengths.
The logarithmic rates of convergence according to grid refinement are computed for each
wave period and these are represented on the lower two axes of Figure 3.3. Here, the top axis
depicts the convergence rate for the three coarsest meshes, 50, 100 and 150 p.p.w., whereas in
the bottom axis the convergence rate for the three finest meshes, 100, 150 and 200 p.p.w. are
shown. The initial logarithmic convergence rates are virtually zero because the wave field is
initialised with almost the same precision in all four simulations. After the initialisation a first
order convergence rate is expected, consistent with the first order upwind scheme used for
the discretisation of the momentum equation. In both cases this is approximately true for the
first couple of wave periods, while it is especially the rate of convergence for the three coars-
est meshes, 50, 100 and 150 p.p.w., that starts to decline after the second wave period. This
indicates that, over time, the increase of the diffusive error, for a simulation with a coarser
resolution, becomes smaller relative to the increase of the diffusive error for a finer resolu-
tion. This could be explained by the fact that the numerical computation of a steeper higher
wave is more diffusive than that of a flatter smaller wave, so, the more the wave is diffused,
the less diffusive the numerical computation becomes.
The time step for each of the resolutions, presented in Figure 3.5, might give an extra in-
dication why the rate of convergence for the coarser resolutions decline at a higher rate than
that of the finer resolutions. The time step of the three finest resolutions, 100, 150, 200 p.p.w.,
become more or less steady after the initial estimated time step of δt = 10−3 is corrected for.
However, the time step of the coarsest resolution, 50 p.p.w., initially remains around the same
value as the time step of the 100 p.p.w. simulation, after which it slowly starts to increase in a
unsteady fashion. This is in contrast to what would be expected from the fixed Courant num-
ber of C o = 0.2 that governs the time step. This increased temporal resolution may have re-
sulted in a lower level of diffusion in the numerical simulation using a resolution of 50 p.p.w.,
leading to the higher rate of decline in the rate of conversion observed in Figure 3.3.
Figure 3.6 presents the diffusive error and rates of convergence for wave propagation sim-
ulations performed with a second order Monotonic Upstream-Centered Scheme for Conser-
vation Laws, otherwise referred to as a MUSCL scheme. Such higher order scheme are usually
used to increase the performance of a numerical model, where the convergence rate of con-
vergence is increased to reduce computational time. Ideally, discretisation of the convective
term of the momentum equation with a second order MUSCL scheme should result in a sec-
ond order convergence rate. However, the description of the free surface is still limited by
the VOF scheme and the time integration is still carried out using a first order explicit Euler
scheme, which will result in a less than second order convergence rate. From the results, pre-
sented in Figure 3.6, it can be observed that the numerical solution has improved in terms
of the absolute diffusive error. However, the increase of the diffusive error over time is not
38 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.5: Time step of the numerical computations, presented as a function, both of which have been normal-
ized by the wave period.

Figure 3.6: Diffusive error, normalized by the wave height, as a function of time, normalized by the wave period.
The convective term of the momentum equation is discretised with a second order MUSCL scheme. The lower
two axes indicate the logarithmic rate of convergence, where (a) represents the convergence rate for 50, 100 and
150 p.p.w. and (b) represents the convergence rate for 100, 150 and 200 p.p.w.

steady, which results in an unsteady convergence rate. This is a problem as the quality of a
numerical solution can not be checked by grid refinement studies. As a result the more stable
first order Upwind scheme shall be used throughout this research.
The phase error is presented as a function of time in Figure 3.7 and Figure 3.8, where the
computations were performed with Courant numbers of C o = 0.2 and C o = 0.01. From the
results of the initial set-up, Figure 3.7, it can be seen that the phase error immediately starts to
increase for the coarsest resolution of 50 p.p.w., while for the three finer resolutions it remains
close to zero for the first three wave periods. While the phase error is positive in the case of
the coarsest resolution, meaning that the wave propagates slower over time, the phase error
becomes negative for the simulations with the finer resolutions, meaning that these waves
propagate faster as time progresses. From the diffusive error it can be concluded that the
amplitude of the waves declines over time in all four cases. If it is presumed that the wave
3.2. C ONVERGENCE STUDY ON THE PROPAGATION OF A FULLY NONLINEAR REGULAR WAVE 39

Figure 3.7: Phase error, normalized by the wave period, as a function of time normalized by the wave period.

Figure 3.8: Phase error, normalized by the wave period, as a function of time normalized by the wave period. The
Courant number used is C o = 0.01.

length remains constant, amplitude dispersion would dictate that the waves would propa-
gate slower as the wave height decreases. As a reference a second set of computations were
performed in which the Courant number was lowered to C o = 0.01, as presented in Figure 3.8.
This change mainly effected the time step since all other variables used in the computation
would remain approximately the same. The second set of computations, presented in Fig-
ure 3.8, shows a steady increase of the phase error for the coarsest resolution, while the phase
error for the finer resolutions remains close to zero. This indicates that the time discretisation
is governing the phase error.
Keeping in mind that the intended purpose of the Navier-Stokes/VOF solver is to prop-
agate waves over no more than a couple of wavelengths, a successful validation against the
fully nonlinear stream function solution has shown that the two-phase Navier-Stokes/VOF
solver is able to provide a respectable reproduction of the single-phase reference solution.
The use of a first order Upwind scheme resulted in a diffusive error in the order of O(10−2 ) for
the finer spatial discretizations of 150 and 200 p.p.w. However, these discretizations do not
ensure that the solution is grid independent, so grid studies are recommended for all follow-
ing cases. Note that no consistent convergence could be obtained with the use of a second
order spatial scheme, therefore, the first order Upwind scheme shall be used throughout the
present research.
40 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.9: Schematic representation of the numerical domain for wave loading on a partially submerged fixed
horizontal cylinder, with l = 6.25m, d = 0.5m, a = 0.2m and D = 0.1m.

3.3. WAVE LOADING ON A PARTIALLY SUBMERGED FIXED HORIZONTAL


CYLINDER

This section elaborates on the ability of the Navier-Stokes/VOF solver to accurately compute
wave forces on a partially submerged fixed horizontal cylinder. The validation cases in this
section were set-up in order to reproduce results from experiments performed by Dixon et al.
(1979).

3.3.1. N UMERICAL SET- UP


The work from Dixon et al. (1979) presents a series of experiments in which vertical and hor-
izontal forces on a fixed horizontal cylinder, induced by regular waves of varying amplitude,
were determined for various levels of cylinder submergence. In their work, they discovered
that the varying buoyancy force of a partially submerged cylinder can play an equally large
role as the inertial force in the resultant vertical force. In some of their cases these two forces
lead to entirely negative vertical forces that acted at twice the wave frequency. In the Mori-
son’s equations, this would result in a negative inertial coefficient, C M , to cope with this they
introduced a varying volume and buoyancy term to the Morison’s equation. A total of ten
cases from Dixon et al. (1979) were simulated using the waveFoam solver, five with a con-
stant relative axis depth and a varying relative wave amplitude and five with a constant rela-
tive wave amplitude and a varying relative axis depth. Here, the relative amplitude and axis
0
depth, are the ratio of the wave amplitude, A, and the axis depth, d , to the cylinder diame-
ter, D. In order to compare the obtained forces, the dimensionless Root Mean Square (RMS)
vertical force was determined according to
s
1 Xn
F r ms = (F /ρg (πD 2 l /4))2 (3.4)
n k=1

where F r ms is the RMS force over one wave period, calculated from the relative vertical force,
which is the total vertical force component divided by the weight of the water displaced by
a totally submerged cylinder. Both theoretically and experimentally obtained RMS vertical
forces from Dixon et al. (1979) are presented in Table 3.3. A constant relative wavelength
of 15.62 was applied for all cases. The relative wave length is defined as the wavelength, λ,
divided by the cylinder diameter, D.
A schematic representation of the numerical domain is presented in Figure 3.9. The di-
3.3. WAVE LOADING ON A PARTIALLY SUBMERGED FIXED HORIZONTAL CYLINDER 41

Figure 3.10: Visualisation of the numerical simulation of the wave loading on a fixed horizontal cylinder. Here,
the iso-contour with α = 0.5 is used to visualise the free surface. The relative wave amplitude and axis depth were
0.5 and 0.0. The spatial discretisation used for these simulation was 150 p.p.w.

mensions, l = 4λ, d = 0.5m, a = 0.25m and D = 0.1m, were applied to comply with the ex-
perimental set-up from Dixon et al. (1979). The mesh, including the cylinder wall, was cre-
ated with hexahedra by defining individual vertices, blocks and arcs, this can be seen from
Figure 3.10, where the numerical simulation is visualised. A refinement area was defined
around the cylinder surface to maintain the unit ratio cell dimensions as much as possible.
This meant that the cells at the cylinder boundary surface were twice as small as the refer-
ence spatial discretisation. It may be noted that the more than ninety percent of the cells
in the computational domain had the reference cell dimensions, which were based on the
p.p.w. This same method was also applied to generate the computational domains for the
simulations in Section 3.4 and 3.5.

3.3.2. A NALYSIS WAVE LOADING


The following subsection will start with the evaluation of two different relaxation zone set-
ups. This is followed by a grid refinement study preformed to evaluate the sensitivity of the
numerical solution to changes in the spatial discretisation. Finally the results from the ten
simulations, with various relative amplitudes and axis depths, proposed in Table 3.3, will be
analysed. The waves2Foam toolbox provides two methods for conducting the wave loading
experiment. The wave field can either be gradually developed over a certain period of time
using a step function, or it can be initialised at the start of the simulation. Both methods
involve implementation of an inlet and an outlet relaxation zone. The inlet relaxation zone
Table 3.3: Experimental and theoretical vertical RMS forces (Dixon et al., 1979).
Relative Relative Theoretical Experimental
amplitude axis depth F r ms F r ms
0.10 0 0.063 0.062
0.20 0 0.126 0.121
0.30 0 0.185 0.172
0.40 0 0.238 0.221
0.50 0 0.284 0.264
0.20 0.10 0.128 0.121
0.20 0 0.126 0.121
0.20 -0.10 0.115 0.112
0.20 -0.20 0.099 0.101
0.20 -0.30 0.074 0.087
42 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.11: Vertical forcing, normalized by the buoyancy force of the completely submerged cylinder, as a func-
tion of time normalized by the wave period. Comparison of a zero current and wave relaxation condition applied
to the zone outlet. The reference solution is produced with a domain that had a length of 20 wavelengths to the
right of the cylinder. The relative wave amplitude and axis depth were 0.5 and 0.0. The spatial discretisation was
150 p.p.w.

can be set up to reproduce various types of waves. The waves produced in the experiment
were said to be linear, however, in general the waves generated by a piston type wave maker
are not perfectly linear, therefore the stream function wave was used to reproduce the wave
profiles. It should be noted that this could result in small differences between the numerically
computed wave profiles and the ones generated in the physical wave tank. Please note that
the analysis of physical experiments is not trivial and there are more uncertainties in terms
of the reproduction of these experiments.
There are two ways the outlet relaxation zone can be set up: either with a zero current
condition, or with a wave theory condition equal to the one applied at the inlet relaxation
zone. In the former, the target solution is a flat water surface condition, resulting in gradual
absorption of the wave. In the latter, the target solution is the undisturbed incoming wave,
here, the wave disturbed by the presence of the cylinder, is relaxed towards the undisturbed
wave condition. If the disturbance of the wave profile is small, the relaxation zone has a
relatively small task since the computed and the target solutions are close to one another.
Figure 3.11 presents the results of a comparison of the between both types of outlet condi-
tions.
As mentioned before, the wave field can be initialised in two ways; either by using a ramp
function to increase the wave amplitude over time, or initialising the full wave field at the
start of the simulation. The former, in which the simulation starts with a flat free surface, is
3.3. WAVE LOADING ON A PARTIALLY SUBMERGED FIXED HORIZONTAL CYLINDER 43

Figure 3.12: Vertical forcing, normalized by the total buoyancy force of the cylinder, as a function of time normal-
ized by the wave period. Comparison between theoretical and experimental data from Dixon et al. (1979) and
numerical simulation. Grid sensitivity analysis where a resolution of 50, 100, 150 and 200 p.p.w. are applied. The
relative wave amplitude and axis depth are 0.5 and 0.0.

preferred when the zero current condition is applied to the outlet relaxation zone, while it is
more desirable to use the latter when the wave condition is applied to the outlet relaxation
zone. To evaluate the effect of the two different relaxation zone set-ups a reference solution
was created. The simulation for the reference solution was performed in a domain of twenty-
two wavelengths long, twenty of which were located behind the cylinder so that the forcing
on the cylinder would not be influenced by reflections from the outlet boundary. The relative
axis depth of the cylinder was zero and a wave, with a relative amplitude of 0.5, was ramped
up over one wave period, T = 0.9962s. This meant that the total wave height was equal to
the cylinder diameter and that the cylinder would be half submerged in still water. The spa-
tial discretisation applied in these simulations was 150 p.p.w. Figure 3.11 depicts the vertical
forcing on the cylinder for a timespan of 15 wave periods. It is seen that the vertical forcing
of the reference solution is fully developed after approximately five wave periods. From Fig-
ure 3.11a, it can be observed that the vertical forcing with the zero current relaxation, starts
out consistently with the reference solution during the ramp up phase and halfway the fifth
wave period the first minor differences start to appear in the form of an overestimation of
the first and second peak. In the following wave periods these peak forces decline towards
a slight underestimation, after which the solution becomes steady around the ninth wave
period. With exception of these minor differences in the peak forces the force profiles gen-
erally show good correspondence. Although the differences with the reference solution are
relatively small, the results presented in Figure 3.11b, match even better with the reference
solution. When the wave is generated from the start of the computation, the vertical forcing
is not merely in better correspondence with the reference solution, also, this method brings
with it the advantage that it only needs about one wave period to reach the steady state con-
dition. This not only saves a lot of computational time, but also requires less data to be stored.
As a result of this the wave condition will be applied throughout this research where suitable.
To verify the grid dependency of the solution, a grid refinement study was performed with
the spatial discretisations presented in Table 3.4. The results of this grid refinement study
are presented in Figure 3.12, where the vertical forcing on the cylinder was compared with
both the theoretical and experimental data from Dixon et al. (1979). It can be observed that,
as the size of the computational cells decrease, the correspondence between the numerical
44 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.13: The absolute amplitudes, of the various frequencies in the forcing signal, as a function of the fre-
quency, normalized by the wave period. These were calculated using a Fast Fourier Transformation analysis.
Comparison between theoretical and experimental data from Dixon et al. (1979) and numerical simulation with
150 p.p.w. The relative wave amplitude and axis depth are 0.5 and 0.0.

solution and the experimental force profile increase. The coarsest resolutions, with a spatial
discretisation of 50 p.p.w., was considered to low to produce the desired level of accuracy
in the description of the forces acting on the cylinder as the the overall forcing was severely
underestimated. The two finest resolutions, 150 and 200 p.p.w., preformed at a much more
acceptable level. As the difference between these two is hardly noticeable, a resolution of 150
p.p.w. was adopted for the remainder of the simulations in this section.
It can be observed that the numerically computed force profile has better correspondence
with the experimental than with the theoretical solution. By performing a Fast Fourier Trans-
formation on the forcing signals, the results of which are presented in Figure 3.13, the good
correspondence between the numerical and the experimental forcing was confirmed. The
mean of the forcing signal, which is presented on the zero frequency, has a good fit for all
data. The amplitudes of the first and second harmonics of the experimental forcing signal
are slightly overestimated by the theoretical values, while it can be observed that the numer-
ical forcing signal gives a better match. Furthermore, while the the higher frequency loading
amplitudes of the experimental forcing signal were relatively small compared to the primary
frequency, it can be seen that these were reproduced by the numerically computed forcing.
Note that the theoretical forcing hardly captures any of the higher order responses correctly.
It may be noted that an even better fit was established for smaller amplitude waves, as can be
seen in Figure 3.14 and 3.15.
The ten simulations can be split into two categories: In the first five simulations the rela-
tive axis depth, d 0 , is equal to 0.0, such that the cylinder is half submerged, while the relative
amplitude of the incident waves is varied from 0.1 to 0.5, the second set of simulations has
a constant relative amplitude, A 0 , of 0.2, while the relative axis depth is varying from +0.1 to
Table 3.4: Spatial discretisation and total number of computational cells of the numerical domain used for the
wave loading cases.
Points per wavelength Cell size, ∆ [m] Computational cells
50 0.03124 6224
100 0.01562 18144
150 0.01041 40456
200 0.00781 71684
3.3. WAVE LOADING ON A PARTIALLY SUBMERGED FIXED HORIZONTAL CYLINDER 45

Figure 3.14: Vertical forcing, normalized by the total buoyancy force of the cylinder, as a function of time nor-
malized by the wave period. Comparison between theoretical and experimental data from Dixon et al. (1979)
and numerical simulation. The relative axis depth, d 0 , is kept constant at 0.0, while the relative amplitude, A 0 , is
varying from 0.1 to 0.5.

−0.3. Note that two of these ten cases, with d 0 = 0.0 and A 0 = 0.2, are the same so only nine
simulations had to be performed.
The RMS force, F r ms , was computed for all the numerical simulations, the result of which
is presented in Table 3.5. Here, a comparison was made with the experimental data from
Dixon et al. (1979) by computing the absolute error expressed and express this as a percent-
age. It can be observed that the numerical solution generally underestimates the experimen-
tal values, whereas the theoretical solution gives a slight overestimation. Although the corre-
lation is in general quite accurate, around 15 percent, the underestimation gets worse for the
simulations with higher waves.
The RMS forces from by Dixon et al. (1979), Equation (3.4), can be interpreted as the av-
erage force over a single wave period, perhaps more interesting are the actual force profiles
and peak loadings presented in Figure 3.14 and 3.15. Here, the vertical force is given as a
function of time and the numerical results are compared with both the analytical and exper-
imental data. The cases with a constant relative depth are depicted in Figure 3.14, where it is
observed that the analytical solution has a good correspondence with the experimental data,
Table 3.5: Comparison of numerically computed RMS forces and experimental data from Dixon et al. (1979),
whereby the deviation between the two is expressed in percentage.
Relative Relative Theoretical Experimental Numerical Error
amplitude axis depth F r ms F r ms F r ms [%]
0.10 0 0.063 0.062 0.054 13
0.20 0 0.126 0.121 0.104 14
0.30 0 0.185 0.172 0.147 15
0.40 0 0.238 0.221 0.180 18
0.50 0 0.284 0.264 0.210 21
0.20 0.10 0.128 0.121 0.105 14
0.20 0 0.126 0.121 0.104 14
0.20 -0.10 0.115 0.112 0.097 14
0.20 -0.20 0.099 0.101 0.094 7
0.20 -0.30 0.074 0.087 0.074 15
46 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.15: Vertical forcing, normalized by the total buoyancy force of the cylinder, as a function of time normal-
ized by the wave period. Comparison between theoretical and experimental data from Dixon et al. (1979) and
numerical simulation. The relative amplitude, A 0 , is constant at 0.2, while the relative axis depth, d 0 , is varying
from +0.1 to −0.3.

especially for the smaller amplitude waves. For lager amplitude waves, as higher harmonics
start playing a more significant role, see Figure 3.13, the theoretical formulated force signal
provides an underestimation of the first force peak, around t /T = 0.1, and an overestimating
of the second force peaks, around t /T = 0.4. Better consensus with the experimental data
is achieved by the numerically computed forcing. The force profiles are in general better
aligned and especially the peak forces from the more extreme cases show more resemblance.
For the cases with a constant wave amplitude and varying submergence level of the cylinder,
presented in Figure 3.15, a similar result is obtained. Good correspondence is observed for
both analytical and numerical results where the incoming wave has a relatively small ampli-
tude. However, if the cylinder is submerged deeper and the waves are nearly over-topping
the cylinder, an extra peak force becomes tangible, which is described better by the numeri-
cal than by the analytical solution.
These findings were also substantiated by the correspondence of the peak forces pre-
sented in Table 3.6. Here, the minimum and maximum of both the experimental and numer-
ical normalized forces and the corresponding absolute errors between the two are portrayed.
With the exception of a couple of outliers, the errors were found to be in the order of around
5 percent. Please note that there is something peculiar about the experimental forcing signal
of the two cases which had the worst fit with the numerical simulations. The experimental
forcing, of the cases with a relative amplitude equal to 0.2 and submergence levels of -0.2
and -0.3, seem to have moved in time. Although it should be mentioned that the depicted
force profiles in this section were obtained by extracting data points from the figures from
Dixon et al. (1979), which may have resulted into very small discrepancies, the shift of these
experimental forcing profiles is not a result of this. In conclusion of the capabilities of the
Navier-Stokes/VOF solver with respect to the accurate computation of wave forces on a partly
submerged cylinder it is considered that the result obtained with the waveFoam solver is of
acceptable quality.
3.4. S URFACE WAVES GENERATED BY THE FORCED OSCILLATION OF A HORIZONTAL CYLINDER47

3.4. S URFACE WAVES GENERATED BY THE FORCED OSCILLATION OF A


HORIZONTAL CYLINDER

This section explores the ability of the Navier-Stokes/VOF solver to accurately simulate fluid-
structure interaction, to be more specific, the wave generation by the forced heave oscilla-
tion of a two-dimensional horizontal cylinder. For these simulations, with a moving body,
the waveDyMFoam solver was used. The validation cases in this section were set up in accor-
dance with experiments performed by Yu and Ursell (1961).

3.4.1. N UMERICAL SET- UP


In their work, Yu and Ursell (1961), present a theoretical formulation for surface wave gen-
eration by the vertical oscillation of a horizontal cylinder. In this theoretical formulation the
cylinder is initially half submerged, after which the cylinder is forced to oscillate vertically in
a channel of infinite length. To substantiate this theory they performed experiments. Here, a
cylinder was forced to oscillate midway a channel with a finite length of 30m. A selection of
the results from their experiments are presented in Table 3.7. It may be noted that, although
wave absorbers were used at both ends of the channel, the reflections in the experiment were
reported to be in the order of 10 to 20 percent. Numerous experiments were performed, all
with different combinations of heave amplitudes and periods, the results of which were com-
pared using a relative amplitude

Wave amplitude at infinity


RA = , (3.5)
Amplitude of motion of cylinder
where, the ’wave amplitude at infinity’ was evaluated as an average of measurements taken
over a distance of 2 to 6 meter from the cylinder.
Figure 3.16 shows a schematic representation of the numerical domain, which dimen-
sions are defined in correspondence with the experimental set-up from Yu and Ursell (1961).
Relaxation zones with a length of one wavelength were implemented at both ends of the two-
dimensional domain, these were given a zero current condition to absorb the waves gener-
ated by the heaving cylinder. The spatial discretisation was once again based on the p.p.w.,
where the wavelength was calculated from the wave period according to the deep-water limit
according to
Table 3.6: Comparison of the peaks of the relative force profiles of numerical solution and experimental data from
Dixon et al. (1979), where the deviation between the two is expressed as a percentage.
Relative Relative Experimental Numerical Error min Error max
amplitude axis depth F mi n F max F mi n F max [%] [%]
0.10 0 -0.100 0.083 -0.093 0.077 7 7
0.20 0 -0.199 0.147 -0.187 0.136 6 7
0.30 0 -0.278 0.188 -0.287 0.177 3 6
0.40 0 -0.387 0.207 -0.386 0.209 0 1
0.50 0 -0.441 0.250 -0.447 0.244 2 2
0.20 0.10 -0.197 0.161 -0.171 0.158 13 2
0.20 0 -0.199 0.147 -0.187 0.136 6 7
0.20 -0.10 -0.186 0.118 -0.196 0.105 6 11
0.20 -0.20 -0.178 0.085 -0.207 0.091 16 8
0.20 -0.30 -0.169 0.041 -0.189 0.025 12 38
48 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.16: Schematic representation of the numerical domain for wave generation by a heaving cylinder, with
d = 0.58m, a = 0.3m and D = 0.1524m.

g 2
λ= T . (3.6)

A ramp function, with the duration of approximately one period, was implemented in the
cosine function describing the oscillating motion of the cylinder to stimulate a smooth ini-
tialisation of the simulation. The dynamic mesh motion solver, waveDyMFoam, was used to
allow the cylinder to move. Here, an area around the cylinder was defined in which numerical
cells were allowed to deform according to the described motion of the cylinder. The defor-
mation area in the dynamic mesh solver is defined as a radius around the object. A second
radius defines a small area around the object in which the computational cells are not allowed
to deform. It may be noted that the deformation area must be larger than the displacement
of the cylinder to prevent failure of the numerical model, as this inevitably happens when the
cylinder moves outside of this area. Numerical failure was experienced to occurs at an earlier
stage due to the high level of cell deformation caused by the compression of the computa-
tional cells between the edge of the body surface and the border of the deformation area.

3.4.2. A NALYSIS WAVE GENERATION


This subsection will first explore the implementation and implications of the relaxation zones,
provided by the waves2Foam toolbox, in combination with a moving mesh. Secondly, a grid
refinement study will be discussed in which the sensitivity of the solution to changes in the
spatial discretisation will be analysed. Finally, the simulations, with various heave ampli-
tudes and periods, proposed in Table 3.7 are evaluated.
To verify that the implementation of the relaxation zones is not significantly influencing
Table 3.7: Relative amplitude for various periods and amplitudes of the heave motion (Yu and Ursell, 1961)
Heave amplitude Heave period Experimental
[m] [s] RA
0.00615 0.92 0.451
0.00615 0.84 0.514
0.00615 0.76 0.543
0.00615 0.69 0.610
0.00615 0.60 0.695
0.00455 0.51 0.810
0.00465 0.46 0.852
3.4. S URFACE WAVES GENERATED BY THE FORCED OSCILLATION OF A HORIZONTAL CYLINDER49

Figure 3.17: Visualisation of the numerical simulation of the wave generation by the forced heave oscillation of
horizontal cylinder. Here, the iso-contour with α = 0.5 is used to visualise the free surface. The heave amplitude
and period were 0.00615m and 0.92s and the applied spatial discretisation was 100 p.p.w.

the solution, a reflection analysis was performed. In this analysis the cylinder was given a
heave amplitude and period of 0.00615m and 0.92s. A visualisation of the numerical simula-
tion is depicted in Figure 3.17, since the motions of the cylinder were small and the generated
waves even smaller one can hardly observe them. The reference solution was developed by
computing the simulation in an extremely long domain with twenty wavelengths to either
sides of the cylinder, in order to prevent possible reflections contaminating the free surface
elevation near the cylinder.
In Figure 3.18, the reference solution, is compared against a simulation performed in a
’short’ domain, with a total length of eight wavelengths and the cylinder placed in the mid-
dle. Figure 3.18a represents the free surface elevation measured with ’wave gauge 1’ located
one wavelength from the cylinder. It can be observed that the amplitude of the wave de-
velops over the first eight wave periods, after which the signal from the reference solution
becomes constant. The solution obtained from the ’short’ domain shows approximately the
same result for the first eight periods, after which discrepancies start to appear. It could be,
that although a linear wave maker is used, some spurious free waves are generated, similar to
what happens with piston type wave makers in physical wave tanks, a phenomenon of which
a comprehensive review is given in the work of Schäffer and Steenberg (2003). This should
however mean, that these spurious waves would also occur in the reference signal, which
does not seem to be the case. This leads one to believe that the generated waves were not
fully absorbed by the relaxation zones and that the disturbances are caused by wave energy
that is reflected back into the numerical domain. These reflected waves, propagating towards
the cylinder, could be interfering with the outwards propagating waves, causing the distur-
bances. Furthermore, a slight change in period of the wave signal can be observed. It may
well be that these both these changes are caused by the reflection of waves from the relaxation
zones, as both changes seem to take effect from the seventh period, which is approximately
the time it takes for the first wave to travel to relaxation zone and back to the wave gauge.
The same result can be observed from Figure 3.18b, where the surface elevation from
’wave gauge 2’, located two wave lengths from the cylinder, is presented. Although logically,
the signal starts to develop one wave period later, it can be seen that the peak in the secondary
wave pattern, observed during the fourteenth wave period from the signal of ’wave gauge 1’,
is witnessed during the thirteenth period from the signal of ’wave gauge 2’. This would make
sense if the changes in the signal were the effect of reflecting waves as it takes the wave one
period less to travel from the cylinder to the relaxation zone and back to ’wave gauge 2’.
A reflection analysis, available as a post-processing tool in waves2Foam, was performed
50 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.18: Surface elevation, normalized by the experimental wave amplitude, as a function of time normal-
ized by the wave period. Comparison of the surface elevation signals from wave gauges at (a) one and (b) two
wavelengths from the cylinder. The reference solution was produced with a simulation with a long domain of 20
wavelengths to the either side of the cylinder. The heave amplitude and period were 0.00615m and 0.92s and the
applied spatial discretisation was 100 p.p.w.

on the three different signals from Figure 3.18 to identify if waves were indeed reflected by
the relaxation zones. The result of this analysis is depicted in Figure 3.19, where the absolute
amplitudes from the various harmonics are presented as a function of the frequency, normal-
ized by the wave period. It can be observed that the zero frequency, or mean amplitude, of
both the ’normal’ and the ’reflected’ signals all had approximately the same non-zero value.
Furthermore, it can be observed that in contrary to the reference solution, the ’reflected’ sig-
nals from the other two simulations contain waves of a certain amplitude. The difference
in amplitude, between ’wave gauge 1’ and ’wave gauge 2’, may be the effect of numerical
diffusion. This effect, although smaller and obviously in opposite direction, can also be ob-
served in the ’normal’ signal. The refection analysis also revealed a small amplitude for the
frequency where f ·T = 7. This could be the indication of the existence of the secondary wave
pattern observed in the surface elevation signal presented in Figure 3.18. The analysis indi-
cated that the amplitude of this response was approximately equal for both the ’normal’ and
the ’reflected’ signals, which may indicated that a higher order wave frequency was travelling
in both directions. Since no reflected waves were observed for the reference simulation, it
was presumed that this was an effect caused by the application of a moving mesh in combi-
nation with the presence of relaxation zones. It may be noted that excellent functionality of
the relaxation zones for static mesh cases was shown in numerous researches, e.g. Jacobsen
et al. (2012), Paulsen et al. (2014a) and Paulsen et al. (2014b). As the main purpose of this
validation case was to verify the accuracy with which the fluid-structure interaction is mod-
3.4. S URFACE WAVES GENERATED BY THE FORCED OSCILLATION OF A HORIZONTAL CYLINDER51

Figure 3.19: The absolute amplitudes from the reflection analysis, performed on the normalized surface elevation
signals from Figure 3.18, as a function of the frequency, normalized by the wave period. Comparison between the,
’normal’ harmonics, travelling away and the, ’reflected’ harmonics, travelling towards the cylinder.

elled, no further time was spend on finding the exact cause of the observed disturbances, but
instead, a method for excluding the disturbance from the measurements was devised.
The most simple solution to exclude the disturbances would be to run all computations
in the extremely long domain in which the reference solution was created, however, to save
computational time and storage space, a small study was performed to find a trade-off be-
tween domain length and the amount of disturbance. In Figure 3.20, the result from Fig-
ure 3.18a, is compared to a computation performed in a ’long’ domain, which was twice as
long as the ’short’ domain. It can be observed from Figure 3.20b, that the solution was able to
fully develop without the interference of reflected waves. Therefore, the ’long’ domain set-up
was applied for all following computations in this section.
The next step would be to evaluate the dependency of the solution to the spatial discreti-
sation of the numerical domain. This was evaluated in a grid refinement study, where grid
resolution of 50, 100, 150 and 200 p.p.w. were compared. The grid resolutions and the total
number of computational cells in the numerical domain can be found in Table 3.8. Just as
in the evaluation of the relaxation zones and domain length, a heave amplitude and period
of 0.00615m and 0.92s were used. In Figure 3.21, the normalized free surface elevations were
presented for duration of ten periods. Taking into account the ramp time of one period and
the period which is necessary for the wave to reach the wave gauge, it takes the wave approx-
imately eight periods to fully develop, as was also observed during the reflection analysis.
Once more it can be observed that a resolution of 50 p.p.w. gives a poor result, not only a sig-
nificant difference in the hight of the progressing wave, but a small phase difference can be
noticed comparing it to the solutions of the finer resolutions. The solution is seen to improve
as the spatial resolution increases and the difference between the two finest resolutions, of
150 and 200 p.p.w., is so small that it cannot be observed from Figure 3.21. The error of the
Table 3.8: Spatial discretisation and total number of computational cells of the numerical domain used for the
grid dependency study, where the heave amplitude and period were 0.00615m and 0.92s.
Points per wavelength Cell size, ∆ [m] Computational cells
50 0.02 16600
100 0.01 65400
150 0.0066 149400
200 0.005 263600
52 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.20: Surface elevation, normalized by the experimental wave amplitude, as a function of time normalized
by the wave period. Comparison of computations performed in a domain with (a) 3 and (b) 6 wavelengths to
either sides of the cylinder. The reference solution was produced with a simulation with a long domain of 20
wavelengths to the either side of the cylinder. The heave amplitude and period were 0.00615m and 0.92s and the
applied spatial discretisation was 100 p.p.w.

three finer meshes were all in the order of O = 101 , so for the sake of the computational time
the 100 p.p.w. was considered adequate for the final computations.
A total of seven numerical simulations were performed where the cylinder was forced
to oscillate with heave amplitudes and periods matching the experimental work by Yu and
Ursell (1961). The results from these computations were presented in Table 3.9, in the form of
relative amplitudes and a deviation error expressed as a percentage. The relative amplitudes
were calculated from the surface elevation signal of a wave gauge located one wavelength
from the cylinder. It may be noted that both the location of the wave gauge, the size of the
domain and the size of the computational cells was different for each case as the wavelength
for the individual cases varied. Furthermore, the time it took to fully develop the wave profile
varied from case to case. This meant that, for each individual case, a period was selected from
which the wave amplitude was extracted. In all cases a the numerical relative amplitude was a
slight underestimation of the experimentally obtained value. The error between the two was
in the order of ten percent. Considering a spatial discretisation of 100 p.p.w., this meant that,
for the case with a heave amplitude and period of 0.00615m and 0.92, the error between the
two numerically computed and experimentally obtained amplitudes was in the order of O =
10−4 , while the spatial discretisation was in the order of O = 10−2 . It was therefore concluded
that the fluid-structure interaction concerning a dynamic mesh problem was successfully
simulated by the collaboration of the waveDyMFoam solver.
3.5. F REE HEAVE DECAY OF A HORIZONTAL CYLINDER 53

Figure 3.21: Surface elevation, normalized by the experimental wave amplitude at infinity, as a function of time
normalized by the wave period. Comparison of a different grid resolutions, with the surface elevation signal from
a wave gauge at one wavelength from the cylinder. A Courant number of C o = 0.02 and resolution of 50, 100, 150
and 200 p.p.w.l. are applied. The cylinder stroke and wave period are 1.23cm and 0.92s respectively.

3.5. F REE HEAVE DECAY OF A HORIZONTAL CYLINDER


This section elaborates on the ability of the Navier-Stokes/VOF solver to accurately simulate
the decay of the heave motion of a free floating two-dimensional horizontal cylinder. Here,
the flow-induced motions of the structure were modelled by a 6-DOF solver, while a dynamic
mesh solver was used to deform the mesh to enable the motion of the boundary surface. The
validation cases in this section are set up in accordance with theoretical work by Maskell and
Ursell (1970) and experimental work by Ito (1977).

3.5.1. N UMERICAL SET- UP


In their work, Maskell and Ursell (1970), present a theoretical formulation for the decay of
heave motion of a free heaving horizontal cylinder. This theory was later validated by experi-
mental work from Ito (1977), where a horizontal cylinder was partially lifted out of the water
surface and then dropped into still water to observe the transient motion of the cylinder.
The experiments performed by Ito (1977) showed excellent correspondence with the theory
from Maskell and Ursell (1970), indicating that this theory was an excellent estimation of the
physics.
Figure 3.22 shows a schematic representation of the numerical domain, with the follow-
ing dimensions; l = 6m, d = 1.24m, a = 0.2m and D = 0.1524m, which were applied to closely
Table 3.9: Comparison of the numerical relative amplitude and the experimental data from Yu and Ursell (1961),
where the deviation error between the two is expressed in percentage.
Heave amplitude Heave period Experimental Numerical Error
[m] [s] RA RA [%]
0.00615 0.92 0.451 0.406 10
0.00615 0.84 0.514 0.464 10
0.00615 0.76 0.543 0.535 1
0.00615 0.69 0.610 0.598 2
0.00615 0.60 0.695 0.664 4
0.00455 0.51 0.810 0.757 7
0.00465 0.46 0.852 0.763 10
54 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.22: Schematic representation of the numerical domain for the decay test of a heaving cylinder, with
l = 6m, d = 1.24m, a = 0.3m and D = 0.1524m.

mimic with the experimental set-up from Ito (1977). The main difference being that, instead
of physical wave absorbers, a relaxation zone with a length of one meter was implemented
at both ends of the domain. The governing length scale for this experiment is the diameter
of the cylinder, D, rather than the wave length, λ. Therefore, the spatial discretisation was
defined according to the number of points per cylinder diameter (p.p.c.d.) was adopted, see
Table 3.10. In correspondence with the experiment, the initial displacement of the cylin-
ders COG, measured from the still water level, was one sixth of the cylinder diameter, or
d 0 = 0.0254m.
The waveDyMFoam solver used to describe the motions of the cylinder, based on the
pressures on the boundary surface, is a combination of the Navier-Stokes/VOF and the 6-
DOF motion solver, this process is described in Subsection 2.3.3. Furthermore, in order to
allow displacement of the boundary surface of the cylinder, a dynamic mesh solver was used.
Here, a fixed area around the structure was defined, in which the numerical cells are allowed
to deform according to the motion of the cylinder. An inner area is also defined, where defor-
mations of the mesh were not allowed, in order to ensure mesh orthogonality and numerical
stability. An outer radius of R out er = d 0 · 10 = 0.25m was defined in order to ensure that the
cylinder could not move outside this boundary which would inevitably result in the termina-
tion of the numerical simulation.
The simulation of a floating structure, where the hydrodynamics and motions of the
structure are decoupled, can introduce negative added mass which in turn may cause nu-
merical instabilities. Dunbar et al. (2015) discussed this phenomenon on the basis of the
two-dimensional heave decay simulation that is also being explored here. The results they
obtains, presented in Figure 3.23, showed clear signs of instabilities in the version of Open-
FOAM® they utilised for their research. They successfully improved the stability of their
solver by the introduction of a strong coupling between the Navier-Stokes/VOF solver and
the 6-DOF solver, a technique also discussed in the work of Seng (2012). The main difference,
Table 3.10: Spatial discretisation and total number of computational cells of the numerical domain used for the
grid dependency study.
Points per cylinder diameter Cell size, ∆ [m] Computational cells
10 0.015 55120
20 0.0072 215904
30 0.005 497316
40 0.0038 857020
3.5. F REE HEAVE DECAY OF A HORIZONTAL CYLINDER 55

Figure 3.23: Solutions for the body displacement, velocity and force from the simulation of free heave decay
of a two-dimensional circular cylinder with a time step size of M t = 0.001s, using the standard OpenFOAM®
interDyMFoam solver (Dunbar et al., 2015).

Figure 3.24: Visualisation of the numerical simulation of the free heave decay of a horizontal cylinder. Here, the
iso-contour with α = 0.5 is used to visualise the free surface. The applied spatial discretisation was 5 p.p.c.d.

besides the lack of relaxation zones, between the solver utilised in the work of Dunbar et al.
(2015), and the waveDyMFoam solver used to simulate the motions of a floating structure in
the present research, is that the waveDyMFoam solver is equipped with an under-relaxation
method. This is a dexterity that is utilised to conserve numerical stability, the implementation
is described briefly in Subsection 2.3.2. This section explores whether the under-relaxation
method can provide numerical stability and decent results in terms of the accurate compu-
tations of the heave decay of a two-dimensional cylinder.

3.5.2. A NALYSIS MOTION DECAY


This subsection will focus on the decay of heave motion simulated with the waveDyMFoam
solver. First the displacement of the cylinder was evaluated and compared against exper-
imental and theoretical data, after which the velocity and force profiles were evaluated in
order to examine the stability of the numerical solution.
A visualisation of the numerical simulation is depicted in Figure 3.24. The vertical dis-
placement of a the COG of cylinder, relative to the still water level, during the decay tests
is presented in Figure 3.25. Here, the vertical displacement is normalized by the initial dis-
placement, d 0 , and presented as a function of time normalized by g /a. The cylinder was
p

dropped from an initial displacement of 0.0254m, with zero velocity. The decay test was per-
formed using four different grid sizes ranging from 10 to 40 p.p.c.d., see Table 3.10 for the
applied discretisations an the total number of computational cells.
It was found that the numerical domain, with a length of 6m, corresponding to the ex-
perimental set-up described in Ito (1977), was sufficiently long enough to exclude reflections
from disturbing the solution of the numerical computation. From Figure 3.25, it can be ob-
56 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

Figure 3.25: Vertical displacement


p of the center of mass, normalized by the initial displacement d 0 , as a function
of time normalized by g /a. Comparison of a different grid resolutions to theoretical and experimental results.
The spatial resolutions 10, 20, 30 and 40 p.p.c.d. were applied.

served that, as the cylinder gains velocity on its first descend, there was hardly any difference
between the solutions of different spatial discretisations, and that the correspondence with
the theoretical solution was excellent. The minor differences between the numerically com-
puted displacement and the theoretical and experimental solutions indicated that, in the nu-
merical domain, the cylinder gained more velocity during its descent. This indicated that
slightly less damping was present in the numerical domain.
The four numerically computed solutions, of different spatial resolution, were seen re-
main in good correspondence with the the theoretical solution as the cylinder moves in the
positive, upward, direction. Small discrepancies between numerical and theoretical result
p
only started to appear around t / (g /r ) = 7, as the cylinder reached the maximum upward
displacement, where a minor overestimation of the peak response can be observed. This
could again be the effect of an underestimation of the damping in the numerical simulation.
Please note that a comparison is being made between a two-phase numerical and experi-
mental solution and a single phase theoretical solution. For a larger part of the first heave
period, no significant difference was observed between the different spatial discretisations,
p
small discrepancies can however be observed around t / (g /r ) = 7. It may be seen as a token
of grid convergence that, the higher the spatial resolution the better the correspondence with
both the theoretical and the experimental value.
p
The first response peak, around t / (g /r ) = 7, was analysed in slightly more detail. Sim-
ilar to Dunbar et al. (2015), the numerical and experimental signals were compared in terms
of amplitude and phase error. The results are presented in Table 3.11, from which can be
observed that the period of the numerically computed transient motion was in close prox-
imity of the experimental value. The error in the amplitude however, was in the order 10 to
20 percent. This meant that, depending on the specific case, the amplitude error was in the
Table 3.11: Comparison between the numerical solutions and experimental data from Ito (1977) of the first peak
in the cylinders vertical displacement profile, where the period and amplitude error were determined.
Points per cylinder diameter Error period [%] Error amplitude [%]
10 1.21 19.5
20 1.59 14.4
30 1.27 12.6
40 1.29 12.1
3.5. F REE HEAVE DECAY OF A HORIZONTAL CYLINDER 57

Figure 3.26: Vertical velocity (a), normalized by the maximum velocity v max , vertical acceleration (b), normalized
by the maximum acceleration, a max , and vertical force (c), normalized p by the buoyancy force of the completely
submerged cylinder. All presented as a function of time normalized by g /a. The spatial resolutions 10, 20, 30
and 40 p.p.c.d. were applied.

order of one tenth to a half of the computational cell size. It shall be noted that the theoreti-
cal and experimental transient motions depicted in Figure 3.25, were obtained by extracting
data points from the figures in Ito (1977). This may have resulted into small discrepancies
in the displacement profiles and consequently in the calculated period and amplitude errors
presented in Table 3.11.
From vertical displacements, presented in Figure 3.25, it seems that the numerical solver
provides an accurate solution to the problem. To examine the reliability of the waveDyM-
Foam in terms of stability, the velocity, acceleration and force profiles of the cylinder were
evaluated, these are presented in Figure 3.26. The velocity of the cylinders COG, normalized
by the maximum velocity is presented, in Figure 3.26a, as a function of time, normalized by
the square root of the gravitational acceleration divided by the radians of the cylinder. No
significant disturbances could be observed in the velocity profiles. Figure 3.26b, depicts the
vertical acceleration of the cylinders COG and the vertical force on the cylinder, normalized
by the buoyancy force of the totally submerged cylinder is presented Figure 3.26c, both of
58 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

these are also presented as a function of the normalized time. The vertical acceleration, ve-
locity and displacement are all governed by the vertical forces on the cylinder, as is described
in Section 2.3. This becomes especially clear when the vertical force in is compared to the
vertical acceleration. The force spikes which can be observed in the solutions from all four
numerical simulations are also observed in the vertical acceleration of the cylinders COG.
The observed force spikes in the velocity profile are relatively small, and smaller even than
the ones observed in the force profile. This may be an effect of the implemented under-relax-
ation method, which uses an ’accelerationRelaxation’ factor to relaxes the acceleration calcu-
lated from the forces on the cylinder, see Subsection 2.3.2 for a more elaborated description
of this technique. It should be noted that the disturbances that can be observed from Fig-
ure 3.26, are by no means as exceedingly large as the ones presented by Dunbar et al. (2015),
as presented in Figure 3.23. This could mean that the stability of OpenFOAM® was improved
by the implementation of the under-relaxation method.
A small investigation was executed to evaluate the influence of the ’accelerationRelax-
ation’ factor on the solution of the flow-induced motions of a floating structure. For this study
five heave decay simulations of the two-dimensional horizontal cylinder were performed,
where the only variation between the cases was the ’accelerationRelaxation’ factor. A spatial
discretisation of 20 p.p.c.d. was used since this was seen to produce descent results in terms
of stability, as seen from the velocity, acceleration and force profiles in Figure 3.26. The sim-
ulations were performed with a ’accelerationRelaxation’ factor of 1.0, 0.7, 0.5, 0.3 and 0.0, the
results of which are presented in Figure 3.27.
From Figure 3.27a, where the vertical displacement of the cylinders COG is presented,
it can be seen that the fully relaxed simulations, i.e. the simulation for which the ’accelera-
tionRelaxation’ factor equals 0.0, the cylinder does not move. The cause for this can be found
in the definition of the under-relaxation method. The acceleration at the second time step,
a 2∗ , is equal to f a 2 + (1 − f )a 1 , which is zero, because the initial acceleration, a 1 , is defined as
zero. Furthermore, can be observed from the figure that, the higher the ’accelerationRelax-
ation’ factor, the better the correspondence between the numerical and experimental result.
From the accelerations and force profiles, presented in Figure 3.27b and 3.27c, it may be seen
that, the higher the ’accelerationRelaxation’ factor, the more severe the disturbances. The
simulation performed without under-relaxation of the acceleration, so where the ’accelera-
tionRelaxation’ factor equals 1.0, the acceleration and force profiles displayed heavy oscilla-
tions similar to those presented in the work of Dunbar et al. (2015). This shows that the inter-
DyMFoam solver in its self shows signs of instability and that this stability may be improved
by using an under-relaxation method. It should be noted that, as observed from the vertical
displacement depicted in Figure 3.27, the implementation of under-relaxation seems to have
a negative effect on the convergence of the solution. Taking into account both the reliabil-
ity and the accuracy of the solution, it was decided that the default ’accelerationRelaxation’
factor of 0.5 will be used for all further simulations in the present research.

3.6. S UMMARY
This chapter explored the capabilities of the Navier-Stokes/VOF solver in a two-dimensional
domain. Wave propagation with and without the presence of an object was considered.
Successful grid convergence on the propagation of a fully nonlinear stream function wave
(Fenton, 1988) has shown that the two-phase Navier-Stokes/VOF solver was able to provide
3.6. S UMMARY 59

Figure 3.27: Vertical displacement of the center of mass (a), normalized by the initial displacement d 0 , verti-
cal acceleration (b), normalized by the maximum acceleration, a max , and vertical force (c), normalized p by the
buoyancy force of the completely submerged cylinder. All presented as a function of time normalized by g /a.
Comparison of different ’accelerationRelaxation’ factors. The spatial resolutions 20 p.p.c.d. was applied.

a respectable reproduction of the single-phase reference solution of a fully nonlinear stream


function wave Fenton (1988). First order convergence was observed in accordance with the
first order Upwind and Euler schemes applied for the spatial and temporal discretisation.
The use of spatial discretisations of both 150 and 200 p.p.w., lead to an acceptable diffusive
error in the order of O(10−2 ). No consistent convergence could be obtained with the use of
a second order spatial scheme, this lead to the decision to apply first order schemes in all
further numerical simulations in the present work.
Wave loading on a partially submerged fixed horizontal cylinder was evaluated using a
theoretical and experimental reference solution from Dixon et al. (1979). Although the com-
parability of the RMS forcing values was frugal, with errors in the order of 10 to 20 percent,
excellent results were obtained for the peak loading. Furthermore, it was shown that the nu-
merical results were a better fit with the experimental reference solution than the theoretical
solution proposed by Dixon et al. (1979).
60 3. WAVE PROPAGATION AND FLUID - STRUCTURE INTERACTION IN 2D

In a numerical reproduction of the work by Yu and Ursell (1961), it was shown that the
numerical model can generate free surface waves with the forced heave motion of cylinder.
It may be noted that, the implementation of the forced oscillating cylinder lead to signifi-
cant reflections from the domain boundaries, due to the limitations of the wave absorption
zones. In spite of this, decent results were obtained for the relative wave amplitude, R A , of
the generated waves, showing that the fluid-structure interaction is simulated correctly by
the numerical model.
Finally, the Navier-Stokes/VOF solver was used to simulate the heave decay of a free float-
ing two-dimensional horizontal cylinder. The numerical simulation showed satisfactory sim-
ilarity with the theoretical work by Maskell and Ursell (1970) and experimental work by Ito
(1977). It should be noted that some irregularities, in the form of singular force spikes, were
observed in the acceleration solution. These are a direct result of the spikes observed in the
force signal, which are calculated from the pressure field. These minor signs of instability
lead to the evaluation of the under-relaxation method implemented in waveDyMFoam. Here
it was shown that an ’accelerationRelaxation’ factor successfully improves the stability of the
numerical simulation.
4
M ESHING TOOLS

4.1. I NTRODUCTION
In this chapter the capabilities of two meshing tools are evaluated that are provided with the
standard distribution of OpenFOAM®. The validation cases discussed in Chapter 3, where
fluid-structure interaction was considered, concerned simple two-dimensional structures.
However, the structures encountered in engineering practise will, more often than not, have
more complex three-dimensional shapes. It would be a tremendous amount of work to con-
struct the computational mesh of such structures with the method used to generate the meshes
in Chapter 3. Here, the meshes were constructed from hexahedra, by defining individual co-
ordinates and connecting those with lines and arcs, see Figure 4.1a. To increase the appli-
cability of the numerical model, simulations with more complex structures should also be
achievable without to much effort.
As mentioned before, the discussed cases in Chapter 3 were two-dimensional while a lot
of practical engineering problems, regarding free surface flows around objects, are three-di-
mensional. This change from two- to three-dimensional meshes, will significantly increase
the amount of computational cells. Since CFD simulations are generally intensive in terms of
computational effort, and the computational cost will only increase as the amount of com-
putational cells rise, this chapter also explores a method for reducing the amount of compu-
tational cells by implementing efficient grading to the spatial discretisations.
Section 4.2 elaborates on the implications of snappyHexMesh, a versatile three-dimen-
sional meshing tool available in OpenFOAM®. The three cases from Chapter 3, concerning
fluid-structure interaction, were used for the evaluation of this meshing tool. The effective-
ness of using multi-grading meshes will be evaluated in Section 4.3. This is a method that
allows grading of the spatial discretisation, in the form of local mesh refinement, to reduce
the amount of computational cells in the numerical domain. Since the multi-grading func-
tion was only recently developed and not available in version 2.3.1 of OpenFOAM®, version
2.4.0 was used to create the basic multi-graded block mesh, after which all subsequent pro-
cesses and computations were performed in version 2.3.1. to uphold the comparability with
the other computations.

61
62 4. M ESHING TOOLS

Figure 4.1: Visualisation of the mesh around the cylinder boundary surface for the mesh generated for the wave
generation case, where the reference spatial discretisation was 100 p.p.w. (a) the reference, or ’normal’ mesh
generation and (b) snappyHexMesh, or ’snappy’, mesh generation.

4.2. S NAPPY H EX M ESH


This section examines the implementation and implications of the mesh generation utility,
snappyHexMesh, supplied with OpenFOAM®. Three cases involving fluid-structure interac-
tion, from Chapter 3, were simulated using computational domains generated with snappy-
HexMesh. The effectiveness of the mesh generation utility was evaluated.
SnappyHexMesh generates two- and three-dimensional meshes containing hexahedra,
and split-hexahedra, from triangulated surface geometries in Stereolithography (STL) format.
SnappyHexMesh is more of a mesh sculptor than a mesh generator, because it requires an al-
ready existing base mesh. The base mesh can, for instance, be a arbitrarily shaped numerical
wave tank and the surface geometry, a horizontal cylinder. SnappyHexMesh approximately
sculpts the cylinder in the base mesh, by iteratively refining the base mesh and morphing the
resulting split-hex mesh to the surface geometry, as presented in Figure 4.1b. The specifica-
tions of the mesh refinement level are flexible and user-defined.
SnappyHexMesh is a more user-friendly method for constructing complex meshes com-
pared to the the ’normal’ method, where the mesh is generated by defining individual ver-
tices, blocks and arcs, as presented in Figure 4.1a. The solutions obtained with the ’normal’
method, see Chapter 3, were used as the reference solution throughout this section.
It may be noted that, by applying a high refinement level to a mesh with the same ref-
erence mesh size, the computational cells near the cylinders surface become significantly
smaller, as illustrated in Figure 4.1. The increase of the overall number of computational cells
in the numerical domain was in the order order of 10 percent, see Table 4.1. It may therefore
Table 4.1: The amount of cells in the numerical domain and the total computational time, normalized by the
duration of the simulation, for both the reference and the ’snappy’ simulations. The spatial discretisations for the
low resolution, (a), were 50 p.p.w. for the wave loading and forced heave cases and 10 p.p.c.d. for the free heave
case. For the high resolution, (b), this was 100 p.p.w and 20 p.p.c.d.
Wave loading Forced heave Free heave
cells [-] time [-] cells [-] time [-] cells [-] time [-]
(a)Reference 4636 30 16600 266 55120 202
Snappy 5396 83 17882 256 56863 293
Increase[%] 16 181 8 -4 3 45
(b)Reference 18144 74 65400 782 215904 1740
Snappy 19482 227 68198 773 223239 1970
Increase[%] 7 206 4 -3 3 13
4.2. S NAPPY H EX M ESH 63

Figure 4.2: Vertical forcing, normalized by the buoyancy force of the totally submerged cylinder, as a function of
time normalized by the wave period. Comparison of normal mesh generation and use of snappyHexMesh. The
spatial discretisations of 50 and 100 p.p.w. were applied and the relative wave amplitude and axis depth are 0.5
and 0.0.

Figure 4.3: Time step, normalized by the wave period, as a function of time normalized by the wave period from
the wave loading case. Comparison of normal mesh generation and use of snappyHexMesh.

be said that snappyHexMesh is not only a more user-friendly meshing tool, but also allows
introduce local mesh refinement. In the following subsection the results summarised in Ta-
ble 4.1 are discussed per case.

4.2.1. WAVE LOADING ON A PARTIALLY SUBMERGED FIXED HORIZONTAL CYLINDER


The effectiveness of the SnappyHexMesh utility was examined with the computation of wave
loading on a horizontal cylinder, see Section 3.3 for a thorough description of the set-up of
this case. The relative wave amplitude and axis depth for this particular case were 0.5 and 0.0.
The simulation was preformed with both a spatial discretisation of 50 and 100 p.p.w.
The theoretical, experimental, reference and ’snappy’ solutions of the vertical forcing
were presented in Figure 4.2. Here, the vertical forcing was normalized by the buoyancy force
of the totally submerged cylinder and presented as a function of time, normalized by the wave
period. It can be observed that, for both spatial discretisations, the similitude between the
numerical and experimental solution was improved for the mesh generated by the snappy-
HexMesh utility. Especially the reproduction of the secondary loading peak was improved. It
was believed that this result can mainly be credited to the smaller cells in the vicinity of the
structure, increasing the accuracy with which the pressures were calculated and integrated
64 4. M ESHING TOOLS

over the cylinder surface.


From Table 4.1, it was found that the amount of computational cells, for the low and high
resolution, had only increased by 16 and 7 percent, while the computational time had in-
creased with 181 and 206 percent respectively. This difference was believed to be caused by
the fact that the computational time is largely governed by the time step. In Figure 4.3, the
time steps of all four computations are presented. The maximum Courant number, fixed at
C o = 0.20, is controlled by the size of the computational cells, velocity and time step, see
equation (3.2). Since the velocity of the flow is expected to remain approximately unchanged
and the smallest computational cell was significantly smaller, for the mesh created with snap-
pyHexMesh, this results in a smaller time step. It may be noted that, due to the smaller time
steps, the total amount of time steps was increased by 130 and 190 percent for the low and
high resolution. This seems to be closely related to the increase in computational time.

4.2.2. S URFACE WAVES GENERATED BY THE FORCED OSCILLATION OF A HORIZONTAL


CYLINDER
For the evaluation of the snappyHexMesh utility with respect to a moving mesh, the wave
generation by a forced oscillating cylinder case was performed. An elaborate description of
the numerical set-up of this case can be found in Section 3.4. A heave amplitude and period
of 0.00615m and 0.92s were used. The simulation was preformed with both 50 and 100 p.p.w.
spatial discretisation.
Figure 4.4 depicts the free surface elevation, normalized by the experimental wave am-
plitude at infinity, as a function of the time, normalized by the heave period. Here, it can
be observed that, for the case with lower resolution of 50 p.p.w., the obtained amplitude was
significantly enhanced by using snappyHexMesh, while the amount of cells in the numer-
ical domain was only increased by 8 percent. This improvement was far less apparent for
the case with higher resolution of 100 p.p.w. The lower improvement in the higher resolu-
tion case, may well be due to the fact the accuracy of the simulation of wave propagation, in
contrast to wave loading, to a smaller degree governed by the smaller cell sizes in the direct
vicinity of the cylinder surface. Instead it is more dependent on the spatial discretisation in
the direction of wave propagation further away from the cylinder and this was more or less
the same for both the simulations, as can be seen from Figure 4.1.
From Table 4.1, it can be seen that the amount of computational cells, for the low and
high resolution, have only increased by 8 and 4 percent and in contrary to the high increase
in computational time observed in the wave loading case, a minor decrease in computational
time was observed. It is at present unclear as to how the decrease in computational time came
about, there are many possible causes of which a relative change in the performance level
of the numerical cluster is just one. As explained in the previous case, the Courant criteria
would dictate a smaller time step with smaller computational cells which would result in
an increased computational time. From the time steps, represented in Figure 4.5, it can be
observed that there is a minor increase in time step, during the initialisation phase, where
the motion of the cylinder is ramped up, after which the time steps converge towards an
equilibrium. This more or less stable signal shows a periodic oscillation which may be caused
by the velocity of the heaving cylinder. Note that there is hardly any difference in the mean
of the time step signals between the snappyHexMesh and the reference solutions of equal
spatial discretisation.
4.2. S NAPPY H EX M ESH 65

Figure 4.4: Surface elevation, normalized by the experimental wave amplitude at infinity, as a function of time
normalized by the wave period. Comparison of the reference solution and the solution using snappyHexMesh.
The surface elevation signal was obtained from a wave gauge located at one wavelength from the cylinder. The
spatial discretisations of 50 and 100 p.p.w. were applied and the heave amplitude and period of the cylinder were
0.00615m and 0.92s.

Figure 4.5: Time step, normalized by the wave period, as a function of time normalized by the wave period from
the forced heaving cylinder case. Comparison of normal mesh generation and use of snappyHexMesh.

4.2.3. F REE HEAVE DECAY OF A HORIZONTAL CYLINDER


The numerical set-up for the heave decay of a free moving horizontal cylinder, was similar to
the set-up described in Section 3.5. The difference being that the mesh was generated with
snappyHexMesh, resulting in a higher resolution in the vicinity of the cylinder surface.
The simulation was preformed with a spatial discretisation of both 10 and 20 p.p.c.d. The
results of the decaying heave motion, from the experimental data and numerical simulation
with ’normal’ and snappyHexMesh set-up, are presented in Figure 4.6. Here, the vertical dis-
placement of the COG of the cylinder is normalized by the initial displacement with respect
to the equilibrium. Although differences were small, it can be observed that, for the both the
low and the high resolution, the result improved by use of snappyHexMesh. This manifests
itself in a smaller amplitude of the transient response of the cylinder, which is closer to the
experimental result.
The amount of computational cells increased by approximately 3 percent for both the
low and high resolution cases. However, the computational time of the lower resolution case
was increased with 45 percent, while the increase for the higher resolution case was only 13
percent. When looked at the amount of computational time steps, an increase of approxi-
66 4. M ESHING TOOLS

Figure 4.6: Vertical displacement


p of the center of mass, normalized by the initial displacement d 0 , as a function
of time normalized by g /a. Comparison of the reference solution and the solution using snappyHexMesh. The
spatial discretisations of 10 and 20 p.p.c.d. were applied.

mately 70 and 30 percent was observed for the low and high resolution cases. This seems to
be related to total computational time.

4.3. M ULTI - MESH GRADING


Using the standard single expansion ratio to describe mesh grading permits only “one-way”
grading within a mesh block. To refine the mesh around the free surface, it is possible to de-
scribe the numerical domain by defining several blocks with one grading per block. However,
this can become challenging as desired numerical domains become more complex. Further-
more, it has been experienced that using this technique to refine the mesh around the free
surface poses problems, when snappyHexMesh is used to include a geometrical surface from
a STL file. The snappyHexMesh utility defines an area around structure in which the mesh
is refined, see Section 4.2, and if this area extends over several mesh blocks, failure of the
meshing process was experienced.
OpenFOAM® version 2.4.0 and newer, include a multi-grading functionality that can sub-
divide a blockmesh in several sections in any given direction. Different grading levels can
then be applied within each division, see Figure 4.7(b). Using this functionality the amount
of computational cells can significantly be reduced, see Table 4.2, while maintaining an iden-
tical spatial resolution in the vicinity of the free surface. The following sections will examine
the implementation of this functionality using the results obtained with snappyHexMesh on
a uniform mesh, see Section 4.2, as a reference solution.
In the following evaluation, the multi-graded mesh was defined in a way that the resolu-
tion around the free surface, a horizontal strip slightly larger than the cylinder diameter, has
Table 4.2: The amount of cells in the numerical domain and the total computational time, normalized by the
duration of the simulation, for both the reference and the ’multi-grading’ simulations. The spatial discretisations
were 100 p.p.w for the wave loading and forced heave cases and 20 p.p.c.d. for the free heave case.
Wave loading Forced heave Free heave
Case
cells [-] time [-] cells [-] time [-] cells [-] time [-]
Reference 19482 227 68198 1329 223239 1970
Grading 12919 208 41893 1255 125077 1205
Decrease[%] 34 8 39 6 44 39
4.3. M ULTI - MESH GRADING 67

Figure 4.7: Visualisation of the mesh around the cylinder boundary surface for the mesh generated for the wave
generation case, where the reference spatial discretisation was 100 p.p.w. (a) the uniform reference mesh and (b)
multi-gradient mesh, or ’grading’.

Figure 4.8: Vertical forcing, normalized by the total buoyancy force of the cylinder, as a function of time normal-
ized by the wave period. Comparison of uniform mesh and multi-graded mesh with a resolution of 100 p.p.w.
around the free-surface. The relative wave amplitude and axis depth are 0.5 and 0.0.

approximately the same spatial discretisation as the uniform reference mesh. In the remain-
der of both the top and bottom part of the computational domain, a grading of 25 percent was
applied, as presented in Figure 4.7b. It may be noted that small deviations can be present in
the vicinity of the cylinder surface, since the refinement area used by the snappyHexMesh
utility might contain various cell sizes in the base mesh and this could lead to discrepancies
in the final mesh.

4.3.1. WAVE LOADING ON A HORIZONTALLY FIXED CYLINDER


The effectiveness of the multi-grading mesh was evaluated using the wave loading case dis-
cussed in the previous section. In this subsection, the multi-graded mesh will be compared
to the reference solution with a resolution of 100 p.p.w.
From Table 4.2, it can be observed that, for the wave loading case, the overall amount of
computational cells was reduced by approximately 34%, while only minor differences were
experienced in the vertical force profile, see Figure 4.8. Minor discrepancies in the verti-
cal forcing profile were observed around the secondary loading peak. This was most likely
caused by the minor differences between the cell sizes in the vicinity of the cylinder, as this
resolution was determined to be an important influence on the accuracy of the computa-
tion, as discussed in Section 3.3. In general the magnitude of the forcing was considered to
68 4. M ESHING TOOLS

Figure 4.9: Surface elevation, normalized by the experimental wave amplitude at infinity, as a function of time
normalized by the wave period. Comparison of uniform mesh and multi-graded mesh with a resolution of 100
p.p.w. around the free-surface. The heave amplitude and period of the cylinder were 0.00615m and 0.92s.

be consistent with the results obtained with the computationally more expensive reference
solution. It may be noted that, although the amount of computational cells was decreased
by 34 percent, the computational time only decreased by 8 percent. This is believed to be
caused due to the fact that the computational time is, as least for a large part, governed by
the Courant criteria, as discussed in the previous section.

4.3.2. S URFACE WAVES GENERATED BY A FORCED HEAVING CYLINDER


The multi-grading mesh was evaluated using the wave generation case previously discussed,
where the resolution of the mesh around the free surface of the multi-graded mesh was ap-
proximately 100 p.p.w., which was consistent with the spatial discretisation of the uniform
mesh of the reference solution.
By implementing the multi-graded mesh the total number of computational cells was re-
duced by approximately 39 percent, see Table 4.2. It is known that the propagation of waves,
in a numerical domain, is highly dependent on the amount of cells in the direction of prop-
agation. Since the grading was applied in the vertical direction of the domain, the amount
of computational cells in the direction of propagation remains equal. From Figure 4.9 it can
be observed that the implementation of the multi-graded mesh did not result in significant
changes to the numerical solution. From Table 4.2, it is also seen that, without significant
discrepancies to the solution, the computational time had decreased by approximately 8 per-
cent.

4.3.3. D ECAY OF HEAVE MOTION OF A FREE HORIZONTAL CYLINDER


A final evaluation of the multi-grading mesh was performed on the case concerning the heave
decay of a freely oscillating horizontal cylinder. The results of the numerical simulation were
presented in Figure 4.10. The resolution of the mesh around the free surface of the multi-
graded mesh was approximately 20 p.p.c.d., which was consistent with the spatial discretisa-
tion of the uniform mesh of the reference solution.
From Table 4.2, it can be seen that the total number of computational cells was reduced by
approximately 44 percent. The slightly larger decrease of the number of computational cells,
can be attributed to the fact that the vertical direction of this case is relatively large compared
to the previous cases, i.e. the efficiency of this method increases as deeper numerical wave
4.4. S UMMARY 69

Figure 4.10: Vertical displacement


p of the center of mass, normalized by the initial displacement d 0 , as a function of
time normalized by g /a. Comparison of uniform mesh and multi-graded mesh with a resolution of 20 p.p.c.d.
around the free-surface.

tanks are considered.


The results, presented in Figure 4.10, show that the vertical displacement of the COG of
the cylinder was approximately identical for both the multi-grading mesh and the uniform
reference mesh. It may be noted that the computational time was decreased by approxi-
mately 39 percent, which is in the same order of magnitude as the decrease of the number of
computational cells.

4.4. S UMMARY
In this chapter the capabilities of two meshing tools, provided with the standard distribution
of OpenFOAM®, were investigated. The validation cases discussed in Chapter 3, where fluid-
structure interaction was considered, were re-simulated using meshes created with these
tools. It was shown that snappyHexMesh, a versatile three-dimensional meshing tool, could
be used to create cylindrical boundary surfaces in a standard block mesh and that this tech-
nique could also be applied on a multi-graded block mesh, generated with a multi-grading
tool available in, the slightly newer, OpenFOAM® version 2.4.0. and higher. The numerical
simulations showed no unexpected instabilities as a result of the use of either of these meth-
ods. However, an increase computational time was observed for the use of snappyHexMesh.
This was attributed to the smaller computational cells in the vicinity of the cylinder, when
compared to the reference mesh, which had a large influence on the time step due to the
Courant criteria applied in this research. Furthermore, it was shown that the amount of com-
putational cells could be significantly reduced by the use of the multi-grading tool, resulting
in smaller computational times without a significant loss of accuracy.
5
F LUID - STRUCTURE INTERACTION IN 3D

5.1. I NTRODUCTION
This chapter explores the capabilities of the Navier-Stokes/VOF solver in a three-dimensional
set-up with respect to fluid-structure interaction. The aim of this chapter is validate the nu-
merical model by comparison to laboratory measurements performed at MARIN.
This chapter is structured as follows; first a short introduction is given on the ’TO2 Float-
ing Wind’ project and the physical model campaign. This is followed by a comparison of mea-
sured and computed decay tests. First the results from a free heave decay test are compared
to results from Dunbar et al. (2015), secondly, the results from a free pitch and roll decay sim-
ulation are compared to experimental data and finally the results from moored heave, pitch
and roll decay tests are evaluated with respect to the results from physical model tests.

5.2. TO2 FLOATING WIND PROJECT


The ’Toegepaste Onderzoek Organisaties’ or TO2-institutes are a group of Dutch research
institutes and universities which are stimulated by the government. In 2015 several TO2-
projects, focussed around various societal challenges, were carried out by these institutes. The
present research is part of a TO2-project which is related to ’safe, clean and efficient energy’.
The participating companies in this so called ’TO2 Floating Wind’ project are the Dutch re-
search institutes; ECN, MARIN, NLR and Deltares. The project has two main task areas, which
are a) motion and control of the floating wind turbine and b) helicopter movements in off-
shore wind farms. Within task a) ECN, MARIN and Deltares focussed on extending their
knowledge and integral approach of simulation and testing of floating wind turbines. Nu-
merical simulation were performed by ECN, to improve the control strategy for the control-
lable pitch blades of the floating wind turbine. MARIN refurbished their old high speed basin,
equipping it with a new wave maker and a state of the art wind field generator to perform ex-
periments with the floating wind turbine model, see Section 5.3. The main focus of Deltares
was to extend the knowledge of numerical simulation of fluid structure interaction with a
floating object. Since the maintenance of offshore wind farms, be it floating or fixed, will
partly be done by helicopters, the knowledge on helicopter movements in and around the
wakes of wind turbines is essential. NLR and ECN joint forces to extend the knowledge in this
field.

71
72 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.1: Schematic representation of the physical wave tank with the floating wind turbine model. The dimen-
sions are: l = 200m, w = 4.0m and d = 4.0m.

5.3. P HYSICAL EXPERIMENTS


As part of the TO2 project, physical experiments with a floating offshore wind turbine were
performed in one of the wave tanks at MARIN. Unless otherwise stated all dimensions and
properties will be presented in model scale, for which a 1:50 Froude scaling was applied. This
section will discuss the floating wind turbine model and the experimental set-up.

5.3.1. P HYSICAL WAVE TANK


MARIN had transformed their old high speed basin into the new concept basin, this trans-
formation included the installation of a new wave-maker and the development of a state-
of-the-art wind field generator. This wind field generator is neither depicted nor discussed in
the present research, as no numerical simulations will be preformed with a generated airflow.
An illustration of the 200 m long wave tank is depicted in Figure 5.1. The width of the
wave tank is 4 m as is also the water depth. The model was positioned approximately in
the center of the wave tank. The original design of the OC5 mooring system had a total of
three mooring lines, one connected to each of the heave plates of the sub-structure of the
floating wind turbine, and spread out evenly over the seabed. Due to the limitations of the
wave tank, MARIN had to design and develop a different mooring system. They opted for
a front and back mooring line and two horizontal linear springs, this would give the system
restoring properties close to those of the original three line system. In order to get the balance
right, an extension arm was constructed. All attachment points, dimensions and properties
of the mooring system can be found in Appendix D. It may be noted that the front mooring
line is slightly longer than the back mooring line, which had the original design length. This
extension was implemented, because an earlier model campaign had shown that the original
line length was considered to be to short for extreme conditions loading conditions, which
could lead to complete suspension of the mooring line.
5.3. P HYSICAL EXPERIMENTS 73

Figure 5.2: Schematic representation of the sub-structure of the floating wind turbine model. (a) left side view,
(b) back view. Description of the various parts: A = transition piece, B = center column, C = offset column and D
= heave plate. Description of locations: I = COG, II = front mooring, III = back mooring, IV = left spring and V =
right spring.

5.3.2. OC5 FLOATING OFFSHORE WIND TURBINE MODEL


The design of the sub-structure of the OC5 FOWT, hereafter referred to as the floater, was
enspired on the semi-submersibles found in the offshore oil and gas industry. This generic
model was created by a worldwide collaboration of research institutes to stimulate the in-
crease in knowledge about FOWT. A schematic representation of the floater is depicted in
Figure 5.2. More detailed description of the floating wind system, including dimensions and
other physical properties like the moment of inertia, draft and position of the COG can be
found in Appendix D. Note that these are the values for the experimental model and that this
includes the floater, tower and nacelle.
The sub-structure of this particular model was a try-floater, meaning it has three main
vertical columns. Heave plates were attached, to the bottom of these, in order to increase
the stability of the floater. The three main columns are connected to a, more slender, center
column via support braces. The turbine tower, with at its top the nacelle and blades, is con-
nected to this center column. As mentioned in the previous subsection, an extension arm
was constructed to allow the back mooring line to be behind the aft of the floater.

5.3.3. E XPERIMENTS PERFORMED AT MARIN


Several tests were performed during the model campaign at MARIN. These tests included:
wave calibration tests, decay tests with and without the mooring system and some wave load-
ing cases with and without wind generation. The cases with an induced wind field will be not
discussed within the scope of this report.
Four different wave cases, presented in Table 5.1, were used for wave calibration tests.
These can be divided into two regular and two irregular cases. For the two regular waves the
wave height and period are given, and for the irregular, JONSWAP, wave spectra the significant
wave height and peek wave period, as well as a peak enhancement factor, γ, are provided.
These properties were based on the wave climate on the East Coast of the United States. The
g
wavelength presented here was determined using the deep water limit, λ = 2π T 2 , note that
the MARIN basin had a depth of 4.0 m, which comes down to 200 meters in real scale.
Decay tests were performed in still water to gain insight in the physical properties of the
floating wind turbine. Pitch and roll decay test were performed without the presence of any
mooring system. Heave, pitch, roll and yaw decay test were performed with the floating wind
turbine with the mooring system installed. Please note that the mooring system was altered
as discussed in Subsection 5.3.1. The decay tests will be discussed and analysed in more
detail in Section 5.4 and 5.5.
The four wave loading cases, as presented in Table 5.1, were part of the model test cam-
paign performed with the moored floating wind turbine. Regretfully the data analysis, of
74 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.3: Schematic representation of the numerical set-up used in the free pitch and roll decay cases. The
length, l , was equal to 4.0m, as was the width of the domain. The hight was 4.624m, of which 4.0m was the water
depth, d , and the rest was for the air phase, a.

these cases, was not performed in time to be included in the present research. To evaluate
the performance of the waveDyMFoam solver, it was therefore opted to perform a proof-of-
concept for the irregular wave loading of a moored floating wind turbine, as a a code-to-
experiment validation was not possible, this will be discussed in Chapter 6. Please note that,
previous to this proof-of-concept, a thorough evaluation of the free and moored decay tests
is performed, the results of which are described in the present chapter.

5.4. F REE HEAVE , PITCH AND ROLL DECAY OF A FLOATING STRUCTURE


This section explores the effectiveness of the waveDyMFoam solver to accurately describe
the heave, pitch and roll motion of the floating offshore wind turbine. In these computations
consider the free floating structure, so, no mooring system is taken into account. First the nu-
merical set-up is described after which the experimental and numerical results are reviewed.

5.4.1. N UMERICAL SET- UP


The numerical domain is presented in a schematic manner in Figure 5.3. It was set-up to
closely resemble the experimental set-up. First a cubically shaped domain was formed, here,
the length, l , and width were equal to 4.0 m while the hight was 4.624 m, of which 4.0 m was
the water depth, d , and the rest was to leave enough space for the motions of the floater and
the presence of the air phase, a. Relaxation zones of two-thirds of a meter were implemented
at the inlet and outlet boundary. Note that for these decay simulations, both of these relax-
ation zones were used to absorb waves generated by the floater. As the physical experiments
Table 5.1: The wave heights and periods, in model scale, for the wave calibration and wave loading tests. The
peak enhancement factor, γ, of the JONSWAP spectrum, was based on the wave climate of the East Coast of the
g
United States. The wave lengths were calculated using the deep water limit, λ = 2π T 2 . Please note that all data is
presented in 1:50 Froude scaling.
Description Type Wave height Period Wave length γ
[m] [s] [m] [-]
Regular 7 regular 0.142 1.7112 4.572 -
Regular 10 regular 0.21 2.0023 6.385 -
Operation JONSWAP 0.142 1.7112 4.572 2.20
Design JONSWAP 0.21 2.0023 6.385 3.00
5.4. F REE HEAVE , PITCH AND ROLL DECAY OF A FLOATING STRUCTURE 75

Figure 5.4: Visualisation of the computational mesh used for the free and moored decay tests. Where (a) repre-
sents the view from the left and (b) the view from behind. Detailed illustration of the computational mesh of the
floater surface boundary is presented in Appendix D.

took place in a 4.0 m wide wave tank with concrete walls, both the lateral boundaries of the
numerical domain were set up as fully reflecting walls.
The domain was then graded in the vertical direction using the multi-grading technique,
as discussed in Section 4.3. The multi-gradient mesh was defined in such a way that a hori-
zontal strip around the free surface, over the full length and width of the domain, had a spatial
discretisation of 5 p.p.c.d. Here, the cylinder diameter, D = 0.24m, of the offset columns was
used as a reference. A visualisation of the computational mesh is presented in Figure 5.4
and a detailed representation of the computational mesh of the floater surface boundary is
provided in Appendix D.
The floater is the only part of the floating wind turbine system that was modelled for
the numerical simulations. Note that the physical properties, such as the draft, mass and
moments of inertia of the structure are for the complete floating wind turbine, so, including
the tower and nacelle. These values, presented in Table 5.2, were determined by MARIN for
their physical model. The floating wind turbine, of which only the floater was modelled, was
defined as a ridged body.
The design drawings of the OC5 model were utilised to generate an STL file of the floater,
after which snappyHexMesh could be utilised to sculpt the surface boundary face in the base
Table 5.2: Properties of the physical floating wind turbine model, as provided by MARIN. These are the physical
properties including the floater, turbine and tower.
Description Symbol Unit Magnitude
Draft (mooring) Tmoor ed [m] 0.40
Draft (no mooring) T f r ee [m] 0.3898
Mass M [kg] 111.664
Roll inertia I xx [kg/m2 ] 49.768
Pitch inertia Iyy [kg/m2 ] 47.556
Yaw inertia I zz [kg/m2 ] 43.814
76 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.5: Visualisation of the numerical simulation of the free pitch decay test of a floating wind turbine. Here,
the iso-contours with α ≥ 0.5 is used to visualise the water. The position of the floater is visualised at time interval
(a) t /T = 0.0, (b) t /T = 0.5, (c) t /T = 1.0 and (d) t /T = 1.5.

mesh, as discussed Section 4.2. It should be noted that, even with snappyHexMesh, the gen-
eration of a reliable surface boundary mesh for such a complicated three-dimensional struc-
ture is not trivial. For this a 4-3-2 refinement was applied, this meant that the cells furthest
from the boundary surface would have half the size of the reference mesh, of 5 p.p.c.d., while
the cells close to the surface boundary would be approximately one twenty-fourth of the size
of the reference mesh. This level of mesh refinement was necessary to make sure that the
mesh did not contain any errors. This resulted in a total number of computational cells in
the domain of almost 1.5 million, as presented in Table 5.3, while the base mesh only had
268128 computational cells.
For all the heave, pitch and roll simulations the initial position of the floater had to be
modelled in the base mesh. This meant that, as OpenFOAM® lacks the ability to rotate or
translate an STL file, the STL file itself had to be customised. This just goes to show that the
set-up of these cases is difficult to say the least, and small errors are easily made. The same
can be said for the set-up and analysis of physical model tests, which makes the comparison
of these two far from trivial.
The free pitch decay test is visualised in Figure 5.5, to give an idea of how such a numerical
simulation looks. The iso-contours with α ≥ 0.5 is used to visualise the water. The position of
the floater is captured at its maximum inclination angle the first 1.5 period of pitch motion.
Small disturbances can be observed around the surface piercing cylinders, it is questionable
Table 5.3: Spatial discretisation and total number of computational cells of the numerical domain used for both
the free and restrained decay cases. Here the spatial discretisation is defined for the cells around the free surface
in the base mesh.
Points per cylinder diameter Cell size, ∆ [m] Computational cells
5 0.048 1487575
5.4. F REE HEAVE , PITCH AND ROLL DECAY OF A FLOATING STRUCTURE 77

Figure 5.6: Draft of the floating wind turbine, i.e. position of the keel with respect to the still water level. Initial
position of the floater was bases on the draft of the experimental model.

whether these are a result the motions of the floater, small errors in the numerical model or a
result of the way the free surface is visualised.

5.4.2. A NALYSIS OF THE FREE DECAY SIMULATIONS


It is likely that there is a difference between the physical and numerical model, be it by a fault
in the numerical calculation of the buoyancy force or by discrepancies between the proper-
ties described in Table 5.2 and the actual test model. This may, among others, cause the draft
of the numerical model to be different from the one presented for the model tests. These
discrepancies may also have an effect on the heave, pitch and roll behaviour of the floater.
Before evaluating the decay tests, the draft of the numerical model was established.
A numerical simulation was performed where the floater was initialised in the proposed
equilibrium position of the physical model, with a draft of 0.3989 m. The draft of the floater
is represented in Figure 5.6. The results of this simulation were quantified in Table 5.4, from
which it can be observed that the there was in fact a difference between the draft of the phys-
ical and numerical model. The draft of the numerical model was found to be 3.64 percent
larger, which amounts to approximately 0.7 m in real scale. This newly determined draft was
applied as the standard for the free decay tests discussed in this section. All other physical
properties, as presented in Table 5.2, remained unchanged. Note that, to some degree, the
change in draft will have an effect on the dynamic response of the floating wind turbine. Fur-
thermore, it should be noted that, for the simulation for the draft evaluation, no restrictions
were defined for the floater. The pitch, roll, and yaw angles experienced during this draft ex-
amination were in the order of O(10−2 ) degrees, which give an indication that the numerically
computed floater was balanced correctly.
Now that the correct numerical draft is known, the validation cases could be simulated
and evaluated, starting with the free heave decay of the floating structure. As no model exper-
iments were performed for this particular case, a code-to-code comparison was performed
with the numerical solutions provided by Dunbar et al. (2015) as a reference. In their research
they provided a ’tightly coupled’ version of the OpenFOAM® interDyMFoam solver and com-
Table 5.4: Comparison of the draft of the physical and the numerical model, where the increase is presented as a
percentage.
Description Draft physical [m] Draft numerical [m] Increase [%]
Without mooring 0.3898 0.404 3.64
78 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.7: Heave displacement of the COG (a), normalized by the initial displacement of 0.02 m. The acceleration
(b), normalized by the maximum acceleration. Both presented as a function of time, normalized by the obtained
heave period. Comparison of the result obtained with the waveDyMFoam solver and results from Dunbar et al.
(2015), both from FAST and their ’tightly coupled’ interDyMFoam solver.

pared it to FAST, a NREL engineering tool. Among others, they performed two heave decay
tests with the OC5 semi-submersible floating wind turbine. It should be noted, that although
no intricate mooring system differences had to be overcome, the comparison of these results
is far from a validation of any sort. It should be regarded more, as an initial step in the eval-
uation of the capabilities of the waveDyMFoam solver. Furthermore, it is a check whether
the under-relaxation technique applied in this research, is as capable of providing a stable
numerical result, as the ’tightly coupled’ model presented by Dunbar et al. (2015).
The results of the first of these two heave decay tests is presented in Figure 5.7. A com-
parison between heave displacement computed by the three different numerical models is
presented in Figure 5.7a, where the displacement is normalized by the initial displacement.
For this first heave decay test this was equal to 0.02 m, or 1 m in real scale, while the sec-
ond heave decay test, presented in Figure 5.8, is performed with an initial displacement six
times as large. From the results of the first decay test it can be observed that the results of the
three different numerical models are in close proximity of each other in terms of the com-
puted heave period. This is quantified by the evaluation of the first response amplitude, as
presented in Table 5.5. It can also be seen that damping of the heave motion, computed by
the waveDyMFoam solver, is much higher than that in the simulations performed by Dunbar
et al. (2015). This becomes even more clear from the displacement of the, more extreme, sec-
ond heave test, presented in Figure 5.8a. Although the period is still in good correspondence
with the one computed with the ’tightly coupled’ interDyMFoam solver, the error in the am-
plitude, computed after one heave period, is almost forty percent. Aside from the differences
5.4. F REE HEAVE , PITCH AND ROLL DECAY OF A FLOATING STRUCTURE 79

Figure 5.8: Heave displacement of the COG (a), normalized by the initial displacement of 0.12 m. The acceleration
(b), normalized by the maximum acceleration. Both presented as a function of time, normalized by the obtained
heave period. Comparison of the result obtained with the waveDyMFoam solver and results from Dunbar et al.
(2015), both from FAST and their ’tightly coupled’ interDyMFoam solver.

present in the amount of damping of the system, which may have numerous causes, it can be
observed that no significant disturbances were observed in the accelerations of the floater,
as presented in Figure 5.7b and 5.8b. This indicates that, although there were differences
between the numerical solutions, the waveDyMFoam solver was able to produce acceptable
results in terms of numerical stability by using the under-relaxation method.
Since the capability of the waveDyMFoam solver was proven to provide adequate solu-
tions to the flow problems concerning the heave decay of a complex three-dimensional struc-
ture, it could now be used in the comparison against experimental results from the decay tests
performed at MARIN.
The experimental and numerical pitch angles of the floating wind turbine are presented
in Figure 5.9a. Here, the pitch angle, normalized by the initial inclination angle of 3.13 de-
grees, is presented as a function of time, normalized by the experimentally determined pitch
period. It should be noted that, in order to isolate the pitch motion, the rotations in roll
and yaw direction were restricted. The model was however allowed to translate in all direc-
Table 5.5: Comparison between the numerical solutions obtained with the waveDyMFoam solver and the ’tightly
coupled’ interDyMFoam solver from Dunbar et al. (2015). Here the absolute period and amplitude error were
determined after one wave period.
Description Initial displacement Error period [%] Error amplitude [%]
Heave 0.02 m 0.7 17.6
Heave 0.12 m 1.5 38.1
80 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.9: Pitch angle (a), normalized by the initial angle of inclination of 3.13 degrees. The angular acceleration
(b), normalized by the maximum angular acceleration. Both presented as a function of time, normalized by the
experimentally obtained pitch period. Comparison of the result from the free decay test performed at MARIN
and numerical solution obtained with the waveDyMFoam solver.

tions making it a 4-DOF system. This technique was also applied to all further pitch and
roll cases in the present research. In general a good correspondence was found between the
experimental and numerical result, as can be seen from the amplitude and period compari-
son presented in Table 5.6. It is however worrisome that the numerically computed response
amplitude at t /T = 0.5 was slightly higher than the initial inclination angle. This increase
indicates that the numerical model of the floating wind turbine is either out of balance, the
illegitimate assumption that the axis of rotation could be fixed, a small error was made in the
set-up of the simulation or that there is some underlying numerical problem. Despite the
small disturbance, the numerically computed result fits quite well with the experimentally
obtained solution. Since the time available for this thesis is limited, it was decided not to
take this small disorder into further consideration. From Figure 5.9, where angular acceler-
ations of the floating wind turbine are displayed, it can be seen that the numerical solution
remains relatively stable up to around t /T = 2, where the solution starts displaying oscilla-
tions, as was observed in the work of Dunbar et al. (2015). This indication of instability was
also observed, in a similar magnitude and at a similar time interval, in the roll accelerations
Table 5.6: Comparison between the numerical solutions and experimental data from MARIN in the decay tests
without mooring. Here the absolute period and amplitude error were determined after one wave period.
Description Initial displacement Error period [%] Error amplitude [%]
Pitch 3.13 deg 1.42 5.58
Roll 3.16 deg 0.49 2.14
5.5. M OTION DECAY OF A FLOATING STRUCTURE WITH RESTRAINTS 81

Figure 5.10: Roll angle (a), normalized by the initial angle of inclination of 3.13 degrees. The angular acceleration
(b), normalized by the maximum angular acceleration. Both presented as a function of time, normalized by the
experimentally obtained roll period. Comparison of the result from the free decay test performed at MARIN and
numerical solution obtained with the waveDyMFoam solver.

presented in Figure 5.10b. This remarkable similarity could have many reasons, and there-
fore needs more elaborated investigation, such an analysis does however not fit within the
scope of this thesis. In general it was shown that the waveDyMFoam solver was able to pro-
vide accurate solutions to the flow problems concerned with the free decay tests of a complex
floating structure, where it should be noted that the execution of a seemingly simple physical
decay test is far from trivial. Also, indications of instability in the waveDyMFoam solver were
observed for the cases that concerned angular displacement.

5.5. M OTION DECAY OF A FLOATING STRUCTURE WITH RESTRAINTS


This section examines the waveDyMFoam solver with respect to the accurate computation
of the heave, pitch and roll motion of the moored floating offshore wind turbine. First the
numerical set-up is described, with special attention to the mooring system, after which the
experimental and numerical results are evaluated.

5.5.1. N UMERICAL SET- UP


The numerical set-up, used for the moored decay cases, is illustrated in Figure 5.11. The
dimensions and spatial discretisation of the numerical domain, Γ, were adopted from the
free decay set-up, described in Section 5.4. The main difference being the presence of the
mooring system. As illustrated in the figure, the dimensions of the mooring system do not
limit the size of the numerical domain. This is because the restraints are not a part of the
82 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.11: Schematic representation of the numerical set-up used in the moored heave, pitch and roll decay
cases. The dimensions of the domain, Γ, are the same as in Figure 5.11.

Figure 5.12: Visualisation of the numerical simulation of the moored heave decay test of a floating wind turbine.
Here, the iso-contours with α ≥ 0.5 is used to visualise the water. The position of the floater is visualised at time
interval (a) t /T = 0.0, (b) t /T = 0.5, (c) t /T = 1.0 and (d) t /T = 1.5.

mesh, but simulated as a force in the attachment point on the floater. It should be noted that
mooring line implementation, provided by Niels Jacobsen (Researcher/advisor at Deltares),
provides a post-processing tool for the visualisation of the catenary mooring lines. This can
be observed from the visualisations of the moored heave decay test presented in Figure 5.12.
The mooring system consists of four restraints: two catenary mooring lines and two hor-
izontal linear springs. The main properties of both the mooring lines and the linear springs
are presented in Table 5.7. The attachment point and anchor coordinates can be found in Ap-
pendix D. Here, the attachment points are denoted for the equilibrium position of the floater.
Please note that these had to be altered according to initial conditions of each individual de-
cay case.
Most of the properties presented in Table 5.7 were the directly copied from the design
drawings of the mooring system for the physical experiment. However, no measurement data
was available of the tension in the horizontal springs. Therefore, an estimated pre-tension
of 0.256N, provided by MARIN, was adopted. Knowing the pre-tension, the length of the
springs at rest could be calculated, according to the formulas for linear springs described in
Subsection 2.3.5. Possible discrepancies between the estimated pre-tension and the actual
5.5. M OTION DECAY OF A FLOATING STRUCTURE WITH RESTRAINTS 83

Figure 5.13: Draft of the floating wind turbine, i.e. position of the keel with respect to the still water level. Initial
position of the floater was bases on the draft of the experimental model.

pre-tension in the physical model tests could lead to irregularities in the results. Please note
that this is just one of many causes for the conceivable variations between the experimental
and numerical results.

5.5.2. A NALYSIS OF THE MOORED DECAY SIMULATIONS


In the analysis of the free decay cases, discussed in Section 5.4, it was already observed that
there were small differences between the numerical and physical properties of the floater.
The addition of a mooring system, especially with the uncertainties in the pre-tension of the
horizontal springs, made it more likely that the floater would have a different equilibrium
position in the numerical computations. Therefore, again a numerical simulation was per-
formed where the floater was initialised in the equilibrium position of the physical model,
with a draft of 0.40 m. The vertical displacement of the floater is presented as a function of
time in Figure 5.13, where the vertical displacement is represented as the draft of the floater.
A couple of observations can be made from the result of the draft evaluation. First, the most
obvious observation, that there was in fact a difference between the draft of the physical and
numerical model. An increase in draft of 2.75 percent, as represented in Table 5.4, was ob-
served, which amounts to approximately 0.011 and 0.55 meter in model and real scale respec-
tively. The increased draft of the numerical model would be applied as the standard for the
Table 5.7: Properties of the restraining system of the physical floating wind turbine model, as provided by MARIN.
Wsub is the submerged weight of the mooring line and k spr i ng is the spring stiffness.
Description Length M sub / k spr i ng
[m] [N/m]
Front mooring 18.38 0.04348
Back mooring 16.71 0.04348
Left spring 1.2114 1.51886
Right spring 1.2114 1.51886

Table 5.8: Comparison of the drafts of the physical and the numerical models, where the increase is presented as
a percentage.
Description Draft physical [m] Draft numerical [m] Increase [%]
Without mooring 0.3898 0.404 3.64
With mooring 0.40 0.411 2.75
84 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.14: Heave displacement of the COG (a), normalized by the initial displacement of 0.0272 m. The accel-
eration (b), normalized by the maximum acceleration. Both presented as a function of time, normalized by the
experimentally obtained heave period. Comparison of the results of the moored decay test performed at MARIN
and the numerical solution obtained with the waveDyMFoam solver.

following computations, while all other physical properties, presented in Table 5.2, remained
unchanged. Please note that neither the position of the floater nor the attachment point
and anchor coordinates were altered, instead, the water level was increased by 0.011 meter.
Secondly, the displacements of the free and moored floating wind turbines, displaced in Fig-
ure 5.13, show that the implementation of the mooring system lowers the natural frequency
of the heave response of the system. It should also be noted that, despite of the implementa-
tion of the restraints, with all the uncertainties that come with it, the observed pitch, roll, and
yaw angles experienced, were in the order of O(10−2 ) degrees. These relatively small rotations
give the indication that the moored floating wind turbine was balanced correctly.
With the correct draft of the numerically simulated model established, the validation
against the physical model tests, performed at MARIN, could commence. First a heave decay
test was performed, where the initial displacement of the floater was a modest 0.027 m. The
displacement and accelerations of the COG of the floater are presented in Figure 5.14. Here,
the heave displacement is normalised by the initial displacement of the floater, while the ac-
celeration is normalised by the maximum acceleration. Both were presented as a function of
time, normalized by the experimental heave period of 17.5 seconds. It can be observed that,
for the first two heave periods, the numerical solution for the heave displacement provides
a good fit with experimental results. The error in period and amplitude is determined after
one heave period, this is presented in Table 5.9. Although the first correspondence of the first
heave period is good, it is observed from the following two periods that the damping of the
numerical system is higher than the damping in the physical experiment. In terms of stabil-
5.5. M OTION DECAY OF A FLOATING STRUCTURE WITH RESTRAINTS 85

Figure 5.15: Pitch angle (a), normalized by the initial angle of inclination of 3.34 degrees. The angular acceleration
(b), normalized by the maximum angular acceleration. Both presented as a function of time, normalized by the
experimentally obtained pitch period. Comparison of the result from the moored decay test performed at MARIN
and numerical solution obtained with the waveDyMFoam solver.

ity, the accelerations, presented in Figure 5.14, only show some minor spikes during the first
half period of the simulation.
The numerically and experimentally obtained displacements of the moored pitch decay
test are presented in Figure 5.15a, where the pitch angle, normalized by the initial pitch dis-
placement of 3.34 degrees, is presented as a function of time, normalized by the experimental
pitch period of 33.1 seconds. To isolate the pitch rotation, the numerical model was restricted
to the pitch rotation, while it was free to translate in the x, y and z direction. As seen from
Figure 5.15 and the calculated period and amplitude error, presented in Table 5.9, the cor-
respondence between the numerical and experimental result was less good than previous
obtained results. It may also be noted that, in contrast to the significantly larger damping
observed for the heave decay test, the damping observed in the numerically computed pitch
motion is much more equivalent to the experimental result. From Figure 5.15b, where the
pitch acceleration is presented, it can be seen that the numerical model, minor spikes put
aside, performed well for over two pitch periods. At around t /T = 2.3 the first significant in-
Table 5.9: Comparison between the numerical solutions and experimental data from MARIN for the moored
decay tests. Here the absolute period and amplitude error were determined after one wave period.
Description Initial displacement Error period [%] Error amplitude [%]
Heave 0.0272 m 1.39 3.04
Pitch 3.34 deg 1.01 26.48
Roll 3.54 deg 3.06 2.76
86 5. F LUID - STRUCTURE INTERACTION IN 3D

Figure 5.16: Roll angle (a), normalized by the initial angle of inclination of 3.54 degrees. The angular acceleration
(b), normalized by the maximum angular acceleration. Both presented as a function of time, normalized by the
experimentally obtained roll period. Comparison of the result from the moored decay test performed at MARIN
and numerical solution obtained with the waveDyMFoam solver.

dications of instability can be observed in the form of violent oscillations in the acceleration
profile. Despite these irregularities, the waveDyMFoam solver managed to produce a decent
result in terms of the pitch decay of a moored floating wind turbine.
Lastly, the roll decay test was performed, the numerical and experimental result of which
are presented in Figure 5.16. Here, a similar phenomenon to all previous pitch and roll simu-
lations can be observed in the acceleration profile presented in Figure 5.16b. It seems as if the
under-relaxation method is capable of maintaining stability during a certain part of the com-
putation after which small signs of instability start appearing. No evident reason was found
for the instigation of the oscillations observed in the acceleration profile, other than that they
are a result of irregularities in the forces acting on the body, as is discussed in Section 3.5. The
roll angle, normalized with the initial inclination angle, was presented in Figure 5.16a, as a
function of time, normalized by the experimentally determined roll period of 34.2 seconds.
Both from the figure and from the analysis of the first response amplitude, presented in Ta-
ble 5.9, it was observed that an excellent correspondence with the experimental result was
achieved. However, again, a small increase in the inclination angle was observed at t /T = 0.5,
the cause of which could be the restriction of the pitch and yaw rotations, to name one. This
phenomenon still needs further investigation. Note that the experimental roll angle shows a
irregularity at negative roll angles, so when the floater is leaning to port side. Although this
abnormality does not seem to have a significant influence on the general progression of the
roll signal, and could just be a measurement error, it shows how non-trivial the evaluation
discussed in this section are. It may therefore be noted that the interpretation of these results
5.6. S UMMARY 87

should be done with a large amount of vigilance.

5.6. S UMMARY
This chapter explored the capabilities of the Navier-Stokes/VOF solver in a three-dimensional
set-up with respect to fluid-structure interaction. The aim of this chapter was to validate the
numerical model by comparison to laboratory measurements performed at MARIN. A short
introduction was given on the ’TO2 Floating Wind’ project and the physical model campaign,
after which the results from the free and moored decay tests with the OC5 floating wind tur-
bine were discussed.
It was shown that snappyHexMesh could be used to generate a sound mesh for the com-
plex three-dimensional sub-structure of the OC5 model. This model was used in the simu-
lation of two free decay tests of various initial displacement. Here, a comparison was made
between the solution obtained with the waveDyMFoam solver and the results presented by
Dunbar et al. (2015). It was shown that the under-relaxation method, implemented in the
waveDyMFoam solver, was equally capable, as the ’tightly coupled’ interDyMfoam solver
from Dunbar et al. (2015), of providing a numerical solution to the intricate flow problems
involved in the decay test. The damping of the heave motion, observed in the waveDyMFoam
solution, was somewhat higher, which became even more apparent in the second decay test,
where the initial displacement of the cylinder was higher.
Numerical simulations of the free pitch and roll decay compared against measurements
from MARIN. Here, excellent results were obtained in therms of the amplitude and period of
the response of the floater. The under-relaxation method was shown to provide a stable solu-
tion for the larger part of the computation, however, at some point during the computation,
oscillations were observed in the angular acceleration signal. These irregularities showed
similarities with the stability issues discussed in Section 3.5.
The numerical and experiment set-up of the moored decay test was discussed. A first in-
dication of the functionality of the mooring line implementation was provided. Considering
the uncertainties involved in the comparison of the numerical and experiments results of the
heave, pitch and roll decay tests, an acceptable result was obtained. The under-relaxation
method was shown to be capable of providing a solution to flow problems involved in these
decay tests, indications of numerical instability were observed in the form of oscillations in
the acceleration profile.
6
P ROOF - OF - CONCEPT
This chapter explores the capabilities of the Navier-Stokes/VOF solver in a three-dimensional
set-up with respect to wave structure interaction. The aim of this chapter is to provide a
proof-of-concept for this type of three-dimensional computations. Please note that it is not
within the scope of this thesis to investigate fatigue loading, impacts on secondary structural
elements or strongly nonlinear phenomenon such as ringing, let alone come up with a inte-
grated design method.
This chapter is structured as follows; first, the capabilities of the potential flow solver,
OceanWave3D, will be discussed, where special attention will be paid to the reproduction
of an irregular wave field and the domain decomposition strategy, and secondly a proof-of-
concept case is of irregular wave loading on a moored floating wind turbine is presented.

6.1. WAVE PROPAGATION USING O CEAN WAVE 3D


Through various validation cases it has been shown that the Navier-Stokes/VOF model can be
used for accurate simulation of waves and wave-structure interaction. The simulations how-
ever, are expensive, both in terms of storage capacity and computation power. In Section 4.3,
a technique was evaluated which reduced the computational time by reducing the number
of computational cells in the CFD domain. This OpenFOAM® functionality has shown to re-
duce the amount of computational time while preserving accuracy of the solution and was
also used successfully in the three-dimensional decay simulations, as discussed in Chapter 5.
Another method for reducing the computational time is to apply a domain decomposi-
tion strategy. Here, the numerical domain is split into an inner domain where the Navier-
Stokes/VOF equations are solved, for the accurate simulation of the wave-structure interac-
tion, and an outer domain where the potential flow equations are used for the time efficient
simulation of realistic sea states. Recently Paulsen et al. (2014b) presented the coupling of
the waveFoam solver with OceanWave3D, a fully nonlinear three-dimensional potential flow
model developed by Engsig-Karup et al. (2009). The implementation of the OceanWave3D is
discussed in Subsection 2.4.2.
This section examines the effectiveness of OceanWave3D and the coupling with Open-
FOAM® in a simulation where the irregular wave field from the MARIN wave calibration tests
is reproduced. The characteristics of the JONSWAP spectrum are presented in Table 5.1.

89
90 6. P ROOF - OF - CONCEPT

Figure 6.1: Schematic representation of the domain of the potential flow solver, Ω, and the waveDyMFoam solver
domain, Γ. WG2 was a wave gauge positioned at location of the center column of the floater and WG1 was located
5.36 m in front of WG2. Note that I+ 21 λ =II+ 12 λ = 5.36m

6.1.1. N UMERICAL SET- UP


The fully nonlinear potential flow solver, OceanWave3D, can use both a wave maker signal
from a piston type wave maker, or a free surface elevation signal from a wave gauge to recreate
a certain time series of propagating waves. It may be noted that the solver can produce both
uni- and multi-directional wave fields, as was presented in Paulsen et al. (2014b). Since no
multi-directional wave propagation was considered and the floating structure did not have to
be modelled, the present evaluation was performed in two-dimensional numerical domains.
As a result of policy regarding the newly installed wave maker in the MARIN concept
basin, unfortunately, no wave maker signals were available. The signals from two wave gauges
could be acquired, which presented the entire time series of the wave calibration tests. As
mentioned above, no structure was present during these tests, which meant that one of the
wave gauges, from here on referred to at wave gauge 2, was located at the position of the
centroid of the center column of the floater. The other wave gauge, from here on referred to
as wave gauge 1, was located 5.36 m in front of wave gauge 2. Since no wave maker signals
were available, the potential flow solver tot reproduce the free surface elevations from a wave
gauge signal. As the waves did not have to be generated at the wave maker, and propagated
approximately 200 m further, to the location of the floater, it may be said that the potential
flow solver was not utilised to its full potential in terms of efficient and accurate wave propa-
gation.
OceanWave3D generates a complete wave spectrum from the free surface elevation signal
from wave gauge 1, which is then used to reproduce the free surface elevation at the location
of the floater, where the free surface elevation is then evaluated by a numerical wave gauge.
Please note that neither in the physical experiments nor the numerical simulations contained
a floating structure as the wave calibration tests were considered.
The potential flow solver determines a cut-off frequency based on the Nyquist frequency,
so the wave spectrum is truncated for the higher frequencies. It should be noted that this
method does not take into account any form of wave reflection that might have been present
in the physical wave tank.
The numerical domain of the coupled OceanWave3D and waveDyMFoam solvers are de-
picted in Figure 6.1, here the smaller CFD domain, Γ, is located inside the larger potential
flow domain, Ω. The coordinates of wave gauge 2 form the center of both the numerical
domains. Two wave generation and absorption zone with a length of two wavelengths were
utilised in the potential flow domain. Wave gauge 1 was located at the end of the wave gen-
eration zone. It may be noted that the coupling relaxation zones in the CFD domain were
relatively short, at only a quarter of a wavelength, which was an effect of the short distance
6.1. WAVE PROPAGATION USING O CEAN WAVE 3D 91

Figure 6.2: Surface elevation, normalized by the significant wave height, as a function of time, normalized by the
peak wave period. (a) experimental wave signals from wave gauge 1 and wave gauge 2, (b) comparison of the
experimental and numerical results for the propagation of a irregular wave measured at wave gauge 2.

between the two wave gauges. The spatial discretisation of both of the numerical domains
is presented in Table 6.1. Note that the resolution of the domain of the potential flow solver
is significantly coarser. This is because it uses a higher order Finite Difference method for
solving the Laplace equation (2.35), and a fourth order Runge-Kutta-scheme for integrating
the dynamic free surface conditions, equation (2.34), which makes it far more accurate than
the lower order Navier-Stokes/VOF solver.

6.1.2. A NALYSIS OF WAVE PROPAGATION USING O CEAN WAVE 3D


The free surface elevation obtained from both the physical wave calibration experiment and
the OceanWave3D computation are presented in Figure 6.2. Here, the surface elevation, nor-
malized by the significant wave height, is presented as a function of time, normalized by the
peak wave period. Figure 6.2a presents a time series of around 500 waves from the physical
wave calibration tests performed at MARIN. These were based on a JONSWAP spectrum with
Table 6.1: Spatial discretisation and total number of computational cells of the two-dimensional numerical do-
main used for the wave propagation cases using the potential flow solver coupled with the waveDyMFoam solver.
Here the spatial discretisation of the waveDyMFoam solver is defined for the cells around the free surface in the
base mesh.
Solver Points per wave- Cell size, ∆ [m] Computational cells
length
OceanWave3D 10 0.64 540
WaveDyMFoam 100 0.064 7800
92 6. P ROOF - OF - CONCEPT

Figure 6.3: Exceedance probability of the crest height elevation, normalized by the significant wave height. Com-
parison of the experimental and OceanWave3D data and a second order Forristall distribution.

a peak enhancement value, γ, equal to 3.00 and a significant wave height and period are 0.21
m and 2.0023 s. A comparison can be made between the experimental wave signal from wave
gauge 2 and the numerically computed free suface presented in ??. On the presented time
scale it is difficult to observe that in general a good correspondence between the numerical
and experimental signals was established. Note that the OceanWave3D simulation starts out
with a flat water condition, a ramp function was utilised to initiate the wave propagation,
since the potential flow solver did not have the ability to initialise a wave field directly.
The crest elevation distribution of both the experimental and numerically evaluated, the
result of which is presented in Figure 6.3. Here, the exceedance probability was presented as
a function of the crest elevation, normalized by the significant wave height. The figure de-
picts the experimental and numerical result as well as the second order Forristall distribution
(Forristall, 1978), which was generated from the characteristics of the JONSWAP spectrum. It
can be observed that for approximately 90 percent of the waves, i.e. the waves up to a crest
elevation of 2 · η max /H s ≈ 1.2, all three exceedance probabilities are in good correspondence.
Especially for the top one percent of the waves it can be seen that OceanWave3D generally
underestimates the crest height measured during the experiment. It may be noted that the
wave heights in the top one percent are significantly higher than what the Forristall distri-
bution suggests, this may however be a result of the relatively low number of waves that was
evaluated here.
Since Figure 6.3 only presents general comparison of the wave spectra, a straight signal-
to-signal evaluation is presented in Figure 6.4. In this analysis a comparison is made, each
time step, between the experimental free surface elevation and the numerical counterpart. In
case of perfect positive correlation this approach would depict a straight line of dots, where
ηOceanW ave3D = η measur ement . As expected from Figure 6.3, the majority of the measure point
,depicted in Figure 6.4, are within the area for which 1 ≤ 2 · η/H s ≥ 1 is true. Please note
that the darker color, as seen from the colorbar in the figure, indicates more measurements
with an approximately equal value. The correlation between the two signals was found to
be acceptable, as majority of the values is centred around the line for which ηOceanW ave3D =
η measur ement . Please note that this analysis does not correct for any differences in phase,
which results in irregularities, like the circular paterns observed. This is deemed to have a
negative influence on the correlation.
From the free surface elevation signal, presented in Figure 6.2b, a substantially high wave
6.1. WAVE PROPAGATION USING O CEAN WAVE 3D 93

Figure 6.4: Square density plot. Comparison of the free surface elevation signals of wave gauge 2 from the exper-
iment and OceanWave3D.

Figure 6.5: Surface elevation, normalized by the significant wave height, as a function of time, normalized by
the peak wave period. Comparison of the measured surface elevation at wave gauge 2 from the experiment,
OceanWave3D and the waveDyMFoam solver.

is seen to pass the position of the structure around t /T p = 430. The computational effort
would be enormous if the entire time series had to be comuted in the waveDyMFoam solver.
Therefore, a coupling between OceanWave3D and waveDyMFoam was established. Here the
time efficient potential flow solver was used to simulate the entire time series, while the, com-
putationally more expensive, CFD computations are only active during short period around
the interesting wave event. There is a one-way coupling of the flow characteristics from the
OceanWave3D solver to the relaxation zones in the CFD domain, more detailed explana-
tion of the coupling is presented in Section 2.4. The free surface elevation of wave gauge
2, from experiment, OceanWave3D and waveDyMFoam, are presented in Figure 6.5. Please
note that no floating structure was present and that the simulations were performed in two-
dimensional numerical domains. It can be observed that some of the higher frequency phe-
nomena are not fully captured by the numerical models. This may partly be caused by the
way OceanWave3D interprets the wave signal data, it generates a uni-directional wave signal
form the experimental wave signal measured at wave gauge 1, as mentioned before, it does
not take into account any waves travelling in the opposite direction. Furthermore it can be
seen that the waveDyMFoam solution is slightly underestimating the wave peaks. Despite the
small discrepancies an acceptable correspondence was established between the free surface
94 6. P ROOF - OF - CONCEPT

Figure 6.6: Visualisation of the computational domain, Γ, used for the irregular wave loading on a moored floating
wind turbine. Detailed illustration of the computational mesh of the floater surface boundary is presented in
Appendix D.

elevations from the measurements and the two numerical simulations.

6.2. M OORED FLOATING STRUCTURE SUBJECTED TO IRREGULAR WAVES


In this section presents an exploratory evaluation of the capabilities of the waveDyMFoam
solver to model the wave loading of the OC5 moored floating wind turbine. The domain de-
composed solver, consisting of the potential flow solver, OceanWave3D, and the waveDyM-
Foam solver, is used to subject the moored floating wind turbine to uni-directional irreg-
ular waves. The aim of this section is to provide a proof-of-concept for this type of three-
dimensional computations.

6.2.1. N UMERICAL SET- UP


The numerical set-up of the domain decomposed solver, used in the present section, is sim-
ilar to the one described in Subsection 6.1.2. The main difference is that the inner domain,
Γ, essentially the set-up of the inner domain, Γ, is equal to the set-up used for the moored
decay tests, as presented in Section 5.5. The sole difference is the length of the domain and
relaxation zones, these are adopted from the two-dimensional set-up described in Subsec-
tion 6.1.2. The inner numerical domain, Γ, is visualised in Figure 6.6, and the spatial dis-
cretisation and the total number of computational cells is presented in Table 6.2. Note that,
with respect to the two-dimensional computations from Subsection 6.1.2, no changes were
applied to the outer domain, Ω. Since no multi-directional wave propagation was required,
there was also no need to use a three-dimensional mesh.
The coupling between OceanWave3D and the Navier-Stokes/VOF solver was fully func-
tional if a static mesh was considered, as shown in Subsection 6.1.2. However, when mesh
motion was introduced, there was a problem with the initialisation of the two-phase flow
field. A pragmatic solution was found to solve this problem: first, the initial time step, t /T p =
425.5, of the waveDyMFoam was created with a zero-current flat water condition, secondly,
Table 6.2: Spatial discretisation and total number of computational cells of the two- and three-dimensional nu-
merical domain of the potential flow solver and the waveDyMFoam solver used for the simulation of a moored
floating wind turbine subjected to irregular waves. Here the spatial discretisation of the waveDyMFoam solver is
defined for the cells around the free surface in the base mesh.
Solver Points per wave- Cell size, ∆ [m] Computational cells
length
OceanWave3D 10 0.64 540
WaveDyMFoam 100 0.064 1600547
6.2. M OORED FLOATING STRUCTURE SUBJECTED TO IRREGULAR WAVES 95

Figure 6.7: Surface elevation, normalized by the significant wave height, as a function of time, normalized by
the peak wave period. Comparison of the measured surface elevation at wave gauge 2 from the experiment,
OceanWave3D and the waveDyMFoam solver.

the OceanWave3D solver was used to compute the flow field until the initial time step of the
waveDyMFoam solver and finally, utilising the end condition of the potential flow solver and
the initial still water condition of the waveDyMFoam solver, the coupled simulation could be
initiated from t /T p = 425.5. Note that, in order for this to work, a ramp function was to be
applied in the waveDyMFoam relaxation zones. This is required, so that the target solution,
which is driven by the potential flow solver, would start from a still water condition.

6.2.2. A NALYSIS MOORED FLOATING STRUCTURE SUBJECTED TO IRREGULAR WAVES


This subsection describes the findings of the simulation of the irregular wave loading of a
moored floating wind turbine in a fully nonlinear wave tank. This proof-of-concept study was
performed to gain insight in the possibilities of the waveDyMFoam solver. The irregular wave
loading conditions were supplied by a fully nonlinear potential flow model, OceanWave3D.
The following is discussed: free surface elevation, motions of the structure and the forces in
the mooring lines.
The free surface elevation from the experimental measurements, the OceanWave3D sim-
ulations and the waveDyMFoam solution are presented in ??. Due to the presence of the
moving floating structure, the interpretation of the free surface, with wave gauge 2, at the lo-
cation of the center column would not work. Therefore wave gauge 2 was shifted 1 meter to
port side, in the waveDyMFoam domain, so it was located in between the floating structure
and the wall and maintained the same x coordinate. It should be noted that, distance from
the floater to the wave gauge was half a meter, and the free surface elevation signal might be
influenced by wave diffraction and radiation.
Due to the implementation of the ramp function, initially, there is little correspondence
between free surface elevation in the waveDyMFoam. It can be seen that the second wave
peak, around t /T p = 428, is overestimated. This is probably a result of the initialisation from
still water, as this was not observed in the result of the two-dimensional coupling presented
in Figure 6.5. Furthermore, an underestimation of the two wave peaks, around t /T p = 429.8
and t /T p = 430.8, are underestimated, even more so than in the two-dimensional simula-
tion. Then, around t /T p = 431, the surface elevation signal from the waveDyMFoam solver
stopped. A numerical error, or floating point error to be more precise, had caused the simu-
lation to stop at this point. Although the waveDyMFoam solver was shown to be susceptible
96 6. P ROOF - OF - CONCEPT

Figure 6.8: Visualisation of the motions of the moored floating wind turbine subjected to irregular waves. Here,
the iso-contours with α ≥ 0.5 is used to visualise the water. The position of the floater is visualised at time interval
(a) t /T = 428.5, (b) t /T = 429, (c) t /T = 429.5, (d) t /T = 430, (e) t /T = 430.5 and (f) t /T = 431.

to numerical instabilities, as presented in Section 3.5 and Chapter 5, this does not directly
imply that the computational error was caused by this instability.
Visualisations of various time steps of the simulations are presented in Figure 6.8. Here,
the background mesh, front and back mooring lines, the floater and the free surface are de-
picted. The flow-induced motions of the floating structure, as well as the resulting stretching
of the front mooring line can be observed from these illustrations. Furthermore, small distur-
bances to the free surface are observed around the surface piercing cylinders. Since all these
effects are of a slightly different spatial scale, they will each be addressed separately.
The most predominant motions of the floating structure subjected to irregular uni-direc-
tional head waves are heave and pitch, since the floating system is symmetrical perpendicu-
lar to the wave direction. Therefore, the heave displacement and pitch angle were presented
in Figure 6.9a and 6.9c. It was shown in Chapter 5, that signs of instability of the numerical
simulation could be observed in the acceleration profiles, therefore, the acceleration in the
heave direction as well as the angular acceleration were also presented, see Figure 6.9b and
6.9d. Both the heave displacement and pitch angle show a smooth profile, however, from the
acceleration profiles it becomes clear that there are signs of numerical instability. These are
observed in the form of oscillations in the acceleration profile, these oscillations are more
intense near the peak oscillations. Furthermore, significant irregularities can be seen in both
the acceleration profiles at the exact same time intervals, just a little after t /T p = 430 and
again around t /T p = 431. The numerical solver recovers from the first irregularity, while
the second one causes numerical failure. This gives an indication that, while the Navier-
6.2. M OORED FLOATING STRUCTURE SUBJECTED TO IRREGULAR WAVES 97

Figure 6.9: Heave displacement of the floater (a), normalized by the maximum displacement. The acceleration
(b), normalized by the maximum acceleration. The pitch angle (c), normalized by the maximum pitch angle.
The angular acceleration (d), normalized by the maximum angular acceleration. All of these are presented as a
function of time, normalized by the peak wave period.
98 6. P ROOF - OF - CONCEPT

Figure 6.10: Visualisation of the catenary mooring lines of the moored floating wind turbine subjected to irregular
waves. Here, the iso-contours with α ≥ 0.5 is used to visualise the water. The position of the floater is visualised
at time interval (a) t /T = 429.5, (b) t /T = 430, (c) t /T = 430.5 and (d) t /T = 431.

Stokes/VOF solver coupled with the 6-DOF motion solver was shown to provide excellent
solutions of intricate fluid-structure interaction problems in Chapter 3, 4 and 5, the numeri-
cal solution is susceptible to numerical errors and the stability issue of the model is not fully
resolved by the implementation of the under-relaxation method.
Figure 6.10 shows some time frames of the numerical simulation, in which the entire
length of the mooring lines is visualised. It can be seen from the figure that the front moor-
ing line is fully suspended in Figure 6.10b and 6.10d. Both of these time frames visualise the
numerical simulation just after a large wave collided with the floating structure. From Fig-
ure 6.11, where the horizontal and vertical mooring forces are presented, it can be seen that
these incidents correspond with large forces in the attachment points of the mooring lines.
It may also be noted that these peak mooring forces coincide with the acceleration peaks ob-
served in Figure 6.9. Please note that it is not argued here, that latter is a result of the former.
What is shown here, is that the mooring line implementation can be utilised to investigate the
mooring system. The accuracy with which this is done remains to be established in thorough
code-to-experiment validation. Also, to further improve the pertinence of the implementa-
tion, the functionalities should be expanded. For example, the fully suspended mooring line
would induce a vertical force in the anchor point, however, the current mooring line imple-
mentation does not provide the calculated anchor loads.
Finally a detailed visualisation of the fluid-structure interaction, during the wave loading
of the floating wind turbine, is presented in Figure 6.12 and 6.13. Here, the free surface is visu-
6.3. S UMMARY 99

Figure 6.11: The horizontal (a), and vertical (b), forces in the attachment points of both the front and the back
catenary mooring line. The forces are normalized by the initial mooring force and presented as a function of time,
normalized by the peak wave period.

alised using the iso-contours where α ≥ 0.5. In these figures highly detailed nonlinear fluid-
structure interactions can be observed, such as wave run-up and wave over-topping of the
structure. Although the accuracy of these computations remains to be validated against ex-
perimental work, they provide a first proof-of-concept that the capabilities of the waveDyM-
Foam solver could provide a contribution to the analysis of complex fluid-structure interac-
tions.

6.3. S UMMARY
This chapter explored the capabilities of the Navier-Stokes/VOF solver in a three-dimensional
set-up with respect to wave structure interaction. The aim of this chapter was to provide a
proof-of-concept for this type of three-dimensional computations. An evaluations of the ca-
pabilities of the potential flow solver, OceanWave3D, was provide. Here it was shown that a
decoupled domain strategy could be used to simulate irregular waves in the waveDyMFoam
solver. This coupling was found to provide a acceptable correspondence with experimen-
tal data. Some limitations were indicated with the coupling when a floating structure was
modelled in the waveDyMFoam solver. An adequate solution was provided for the one-way
decomposed model, which allowed for the simulation of a moored floating wind turbine sub-
jected to irregular uni-directional waves. The analysis of this simulation showed that: a) the
numerical solution was susceptible to numerical instabilities in the form of oscillations in the
acceleration profile, b) the mooring line implementation provided functional restraints for
the floating structure, and c) the waveDyMFoam solver is capable of providing detailed de-
100 6. P ROOF - OF - CONCEPT

Figure 6.12: Visualisation of the motions of the moored floating wind turbine subjected to irregular waves. Here,
the iso-contours with α ≥ 0.5 is used to visualise the water. The position of the floater is visualised at time interval
(a) t /T = 430.5, (b) t /T = 430.6 and (c) t /T = 430.7.

scription of the complex fluid-structure interactions involved in the wave loading of a float-
ing structure. Please note that, although a proof-of-concept was provided, the merit of the
numerical model remains to be evaluated with further validation.
6.3. S UMMARY 101

Figure 6.13: Visualisation of the motions of the moored floating wind turbine subjected to irregular waves. Here,
the iso-contours with α ≥ 0.5 is used to visualise the water. The position of the floater is visualised at time interval
(a) t /T = 430.8, (b) t /T = 430.9 and (c) t /T = 431.0.
7
C ONCLUSIONS AND RECOMMENDATIONS
In this thesis, state-of-the-art numerical models were used to investigate the propagation of
free surface waves and their interaction with fixed and moving structures in two- and three-
dimensional set-ups. For the investigation, two fully nonlinear numerical models were ap-
plied. The fully nonlinear Navier-Stokes/VOF solver, provided as part of the open-source
CFD-toolbox OpenFOAM®, version 2.3.1., which was extended with the implementation of
the wave generation and absorption toolbox, waves2Foam, developed by Jacobsen et al. (2012).
This solver, referred to as waveFoam, was used for validation cases with a static mesh. For
cases where structure motion was demanded, this solver was extended with a 6-DOF motion
solver, to form the waveDyMFoam solver. In an attempt to increase the spatial and temporal
reach of the model a domain decomposition strategy was applied, where the fully nonlinear
Navier-Stokes/VOF solver was coupled with a fully nonlinear potential flow solver, Ocean-
Wave3D, developed by Engsig-Karup et al. (2009). Please note that the fully parallel one-way
domain coupling was established and fully validated by (Paulsen, 2013) and (Paulsen et al.,
2014b). The numerical models applied in this research are open-source and in principle
freely available as part of the waves2Foam framework. It should be noted that, the moor-
ing line implementation, provided by Niels Jacobsen (Deltares), has not been published at
present nor is the implementation publicly available.
The structure of this final chapter is as follows: first, the presented results will be dis-
cussed, secondly, the main conclusions are summarised, thirdly, a few further developments
of the numerical models are suggested and finally recommendations are made for possible
future research.

7.1. D ISCUSSION
The results from a series of validation cases were presented throughout Chapter 3, 4 and
5. Furthermore, a proof-of-concept simulation was provided in Chapter 6. The first chap-
ters concerned an extensive exploration of the capabilities of the Navier-Stokes/VOF solver
with respect to wave propagation and wave structure interaction in a two-dimensional set-
up. Chapter 4 presented the evaluation of two meshing tools provided by the OpenFOAM®
toolbox, which were then utilised to accommodate the investigation of fluid-structure inter-
action in a series of decay tests with a three-dimensional floating wind turbine model. This
research was concluded with a proof-of-concept case, discussed in Chapter 6, involving the
modelling of a three-dimensional moored floating wind turbine subjected to irregular uni-

103
104 7. C ONCLUSIONS AND RECOMMENDATIONS

directional waves. All of these numerical computations were performed in an attempt to


realise the main research objective:

"To establish a well validated fully nonlinear numerical wave tank for the simulation
of complex fluid-structure interaction of moored floating offshore structures."

This section provides a discussion of the results presented in Chapter 3, 4, 5 and 6, in order
to establish the success and limitations of the presented work related to the fulfilment of this
main objective.

7.1.1. A CONVERGING SOLUTION


If this Navier-Stokes-VOF solver is to become a supplement or even a substitute to physical
model testing, one has to be able to count on the solution. Therefore, the solution of the
model should improve as the spatial or temporal resolution increases. A grid convergence
study was successfully performed on the propagation of a fully nonlinear stream function
wave, as described in Section 3.2. The convergence rate was verified for both a finer and a
coarser set of spatial discretisations. An approximately first order convergence rate, corre-
sponding with the first order Upwind and Euler schemes used for the spatial and temporal
discretisation, was observed for the first couple of wave periods. However, the convergence
rate was shown to decrease over time and was lower for the coarser set of spatial discretisa-
tions. In order to increase the computational efficiency an attempt was made to use a second
order spatial scheme, while the results from these simulations showed an improved solution
at some time intervals, no reliable convergence rate could be established. As the present
research would benefit more from a stable result, the choice was made to use the first order
schemes for all the numerical simulations. The various grid sensitivity analysis performed on
all individual validation cases each showed a reliable result, justifying the choice. It should
be noted that increasing computational efficiency is an important issue for these computa-
tionally expensive models and that it would be beneficial to perform more detailed studies
on higher order discretisation schemes.

7.1.2. P ROPAGATION OF A FULLY NONLINEAR WAVE


To be able to investigate the impact of nonlinear waves on floating structures one has to pro-
vide an accurate description of propagating nonlinear waves in the first place. In Section 3.2,
numerical computations with the two-phase Navier-Stokes/VOF solver were used to simu-
late the propagation of the single-phase fully nonlinear stream function wave (Fenton, 1988).
Here, the diffusive nature of these types of models, based on the VOF method, became clear.
Although an excellent resolution was obtained for the initialisation of the flow field, the dif-
fusivity was shown to increase rapidly over the first wave period, during which the air-water
interface was smeared. The description of the free surface by means of a VOF method there-
fore remains an area which is in need of improvement.
In spite of the smearing of the interface during this first wave period, the description of
the fully nonlinear stream function wave was reasonably good in terms of diffusivity, consid-
ering it was a numerical two-phase reproduction of a single-phase flow. An acceptable error
in the order of O(10−2 ) was observed for spatial discretisations of 150 and 200 points per
wavelength. The observations made regarding the phase error of the propagating wave was
reason for concern. The time step, determined by the maximum Courant number, seemed to
be governing the phase error of the wave. A high Courant number, resulting in a larger time
7.1. D ISCUSSION 105

step, lead to a negative phase lag, indicating the unexpected phenomenon that the wave was
moving faster as it got smaller. This is a reason for concern and it would be wise to perform a
more elaborated sensitivity analysis on this subject matter. Considering the time attributed
to this thesis, the remainder of the research questions to be answered, and primarily due to
the fact that an acceptable result was obtained for the intended propagation length of one
to two wavelengths, a more elaborated sensitivity analysis was considered unjustified within
the scope of this thesis.

7.1.3. WAVE LOADS ON A STRUCTURE


The computation of wave loads on a structure was validated on the basis of theoretical and
experimental work by Dixon et al. (1979). The numerical computations lead to accurate re-
sults in terms of peak loading. The Root Mean Square vertical force provided less promising
results, the importance of this force remains questionable at this point, because the numeri-
cally computed loading cycle has a considerably better correspondence with the experimen-
tal than the theoretical values. It should be noted that, as a simplification, the viscous bound-
ary layers on all solid walls were neglected. Despite this simplification the computed wave
loads on the fixed structure were considered to be highly accurate. However, with respect of
the main objective of the thesis, it is clear that thorough validation of wave loading on float-
ing structures is still required. Regretfully, this did not fit within the scope of this research in
terms of time and resources.
The heave decay of a horizontal cylinder was simulated in Section 3.5. The presented
absolute forcing on the cylinder surface showed irregularities, while stable solutions were
provided for the motion and velocity of the cylinder. The observed force spikes may have
been a result of the interpolation of the pressures, which makes it an interpretation prob-
lem, or the force spikes could be a result of errors in the pressure field, making it a numerical
problem. Either way, the disturbances were nowhere near the extreme velocity and force os-
cillations observed in Dunbar et al. (2015), where the instability of the standard OpenFOAM®
interDyMFoam solver was shown. It should be noted that this research probably used a dif-
ferent version of OpenFOAM® and did not implement the waves2Foam toolbox. Since no
instability problems were observed for the various cases, it was concluded that no further
attention would be paid to the force spikes during this thesis.

7.1.4. M OTIONS OF A STRUCTURE


For the simulation of the motions of a floating structure, the model relies on the 6-DOF mo-
tion solver and dynamic mesh handling of the standard OpenFOAM® interDyMFoam solver.
This model was extended with the waves2Foam toolbox in order to implement the generation
and absorption of waves. In Section 3.4, this model was utilised to describe the generation
of waves by the forced heave oscillation of a horizontal cylinder. The results of these two-
dimensional numerical simulations proved to be very sensitive to changes in the domain set-
up. Although eventually an acceptable correspondence with the experimental result from Yu
and Ursell (1961) was achieved, initially, reflections from both of the outer boundaries, in the
order of O(101 ), were disturbing the flow pattern. This meant that the relaxation zones, which
were shown to be perfectly capable of handling wave generation and absorption by Jacobsen
et al. (2012), were not working properly for this particular dynamic mesh set-up. This phe-
nomenon was not observed to effect the results of rest of the dynamic mesh cases involving
free moving bodies. No further time was spend on finding the cause of this problem, as it was
106 7. C ONCLUSIONS AND RECOMMENDATIONS

solely established for this forced motion case and the main goal, to prove capability of the
model to accurately model the fluid structure interaction, was accomplished.
The simulation of the free decay of a horizontal cylinder was presented in Section 3.5.
Here, acceptable correspondence with the theoretical work by Maskell and Ursell (1970) and
experimental work by Ito (1977) was achieved. It should be noted that some irregularities,
in the form of singular force spikes, were observed in the acceleration solution. These are a
direct result of the spikes observed in the force signal, which are calculated from the pres-
sure field. These minor signs of instability lead to the evaluation of the under-relaxation
method implemented in waveDyMFoam. Here it was shown that an ’accelerationRelaxation’
factor, which was provided to relax from the forcing determined accelerations, successfully
improved the stability of the numerical simulation.
After the successful completion of these two-dimensional simulations, the waveDyM-
Foam solver was used for the decay tests of a three-dimensional floating wind turbine support
structure. The free and restrained decay cases were presented in Section 5.4 and 5.5. Espe-
cially the free decay tests provided good similarity with the measurements from the physical
model tests in terms of amplitude and period. Similar accuracy was obtained compared to
the results of the two-dimensional free decay test with the horizontal cylinder.
The presented results of numerical computations of the restrained decay tests showed
minor differences with the measurements from the physical experiments. The results of the
heave decay showed that the numerical model had a little more damping than the physical
model. This could be due to the differences between the numerical representation of the
mooring system and the physical one, due to the estimation that was made for the pretension
in the linear springs, the horizontal restraints of the system, or because the numerical model
actually computed more damping in the heave direction. Since the actual pretension in the
horizontal springs is unknown and no free heave decay tests were done with the model it is
difficult to conclude on this matter.
A comparison against numerical computations of two free heave decay tests, presented
by Dunbar et al. (2015), did show that damping in the waveDyMFoam solver was significantly
higher than in their ’tightly coupled’ interDyMFoam solver. This may indicate that the under-
relaxation method, applied in the present research, introduces extra damping, although it
should be pointed out that there are many uncertainties in this code-to-code comparison.
The differences observed between the numerical and physical results for the restrained
pitch and roll simulations give an indication that there could be something wrong with the
definition of the numerical mooring system. As mentioned, there were some uncertainties
regarding the set-up of the physical mooring system, this makes it difficult to distinguish
whether the errors are caused by the uncertainties in the set-up of the moored floating wind
turbine or that the numerical computation of the governing physics is slightly inaccurate.
Furthermore, it should be noted that indications of numerical instability were observed in
the form of oscillations in the acceleration profile. These indications were supported by the
observations made in the proof-of-concept simulation.

7.1.5. A DVANCED MESHING TOOLS


In Chapter 4, the capabilities of two advanced meshing tools, provided with the standard
distribution of OpenFOAM®, were investigated. It was report that the versatile three-dimen-
sional meshing tool, snappyHexMesh, was an effective tool for sculpting a structure from a
standard block mesh. The method was used to recreate the two-dimensional numerical sim-
7.1. D ISCUSSION 107

ulations from Chapter 3 and proved capable of reproducing an improved level of accuracy in
the solution. The way the mesh generation was set-up lead to smaller computational cells
in the vicinity of the structure. This made for an unfair comparison, because on the one
hand the accuracy of the solution increased, while on the other hand the computational time
increased. More importantly however, the simulations did not show numerical instabilities
due to the implementation of a snappyHexMesh domain and it would therefore be a proper
starting point for the meshing of more complex three-dimensional structures.
In order to lower the computational effort of the model a multi-grading tool, from a newer
version of OpenFOAM®, was utilised to generate mesh with a high cell concentration around
the free surface. Provided the cells in the vicinity of the free surface in the graded meshes had
the same aspect ratio as the uniform reference meshes, a very high similarity was achieved in
the accuracy of the result. The result in terms of time saving did regretfully not prove to be as
consistent. Although a decrease was observed in all cases, just in one of the cases, this was
proportional to the amount of decrease in computational cells.
Both methods were used successfully in Chapter 5, where they were used for the genera-
tion of the numerical reproduction of the physical model experiments of the floating wind
turbine. SnappyHexMesh provided an excellent representation of the substructure of the
floating wind turbine with small computational cells in the vicinity of the structure, while
the overall amount of computational cells was reduced by concentrating the mesh around
the free surface.

7.1.6. F ULLY NONLINEAR NUMERICAL WAVE TANK


Chapter 6 explored the capabilities of the Navier-Stokes/VOF solver in a three-dimensional
set-up with respect to wave structure interaction. Here, a proof-of-concept simulation was
evaluated on the irregular wave loading of a moored floating wind turbine. A decoupled do-
main strategy was shown to be capable of providing irregular wave boundary conditions to
the Navier-Stokes/VOF solver. The potential flow solver, OceanWave3D, was seen to pro-
vide efficient and accurate solutions for the propagation of irregular uni-directional waves.
It should be noted that there were some limitations with the coupling of OceanWave3D and
the waveDyMFoam solver when a moving floating structure has to be modelled, however, an
acceptable solution was provided.
The evaluation of the irregular wave loading of the moored floating wind turbine con-
firmed that, although the under-relaxation method implemented in the waveDyMFoam solver
was shown to provide adequate solutions to intricate flow problems, the numerical solution
was susceptible to numerical instabilities in the form of oscillations in the acceleration pro-
file. It should be noted that in the extreme case these irregularities may lead to break-down
of the numerical computation. In order to use this numerical wave tank as a supplement or
even substitute to physical wave tanks, improvement of the stability of this numerical model
is therefore still a top priority. The mooring line implementation was proven to be func-
tional, although the capabilities of the implementation could be expanded on some aspects
and a full validation of the system is still to be provided. The proof-of-concept simulation
also provided detailed descriptions of the complex fluid-structure interactions, however, full
validation of nonlinear phenomenon such as local impact loads and wave run-up should be
performed to confirm the accuracy of the model.
108 7. C ONCLUSIONS AND RECOMMENDATIONS

7.2. C ONCLUSIONS
In regard to the main objective and the research questions, proposed in Section 1.4, it can be
concluded that:

• A thoroughly validated fully nonlinear solver was presented for the accurate computa-
tion of wave-structure interaction.
• The implementation of a multi-gradient grid can improve computational efficiency
without significant loss of numerical accuracy.
• The waveFoam solver is capable of providing an accurate description of the wave load-
ing on a fixed structure.
• The waveDyMFoam solver can be utilised for the accurate computation of free decay
tests of both two- and three-dimensional floating structures.
• The waveDyMFoam solver is of providing acceptable solutions for the moored decay
tests of a three-dimensional floating structure.
• A functioning restraint system was implemented, which can be used to simulate the
mooring system of a floating structure.
• The domain decomposition strategy can be used to efficiently simulate realistic sea
states in the computational domain of the waveDyMFoam solver.
• The waveDyMFoam solver can be used for the simulation of a moored floating struc-
ture in realistic sea states.
• The numerical stability provided by the under-relaxation method has its limitations
and further improvement of the stability of the waveDyMFoam solver is required.

7.3. F URTHER DEVELOPMENTS


As part of this thesis, a fully nonlinear Navier-Stokes/VOF type numerical model was vali-
dated and verified with respect to wave propagation and wave-structure interaction. Even
though the model has been shown capable of computing highly detailed fluid-structure in-
teractions, including dynamic motion response of a ridged structure, further development is
still necessary. Possible extensions could include:

• Improved stability waveDyMFoam solver: Numerous simulations were preformed in


this research that showed the capabilities of the waveDyMFoam solver, some of the
simulations were seen to be susceptible to numerical instability. The under-relaxation
method which was used here to increase stability was observed to have limitations,
as fluid-structure interaction became more complex and flows more violent. Efforts
should be made in further developments of the numerical models discussed in the
present thesis to increase the numerical stability. Dunbar et al. (2015) provided a differ-
ent approach to increase the numerical stability of a similar numerical model, a thor-
ough code-to-code verification might provide new insights in the matter.
• Improved performance Navier-Stokes/VOF solver: A first order convergence rate of
the Navier-Stokes/VOF solver, consistent with the first order spatial and temporal dis-
cretisation, was provided in this research. During a brief investigation of the effects of
using a second order spatial scheme no stable solution was found. As the applied CFD
model supports various spatial discretisation methods, a thorough validation could be
beneficial to the computational performance.
7.4. F UTURE RESEARCH 109

• Improved turbulence modelling: In all the simulations performed with the Navier-
Stokes/VOF solver, the viscous boundaries were neglected by applying a slip conditions
to solid wall boundaries. As the in most cases the computations were in good corre-
spondence with their experimental counterparts, it was concluded that this approach
was an effective method to simplify the flow problem. A logical next step is to include
turbulence models to improve the wall shear description.
• Improved mooring line description: The mooring line implementation, provided by
Niels Jacobsen (Deltares), was based on a simple catenary type mooring line, which
was not influenced by waves or current. Firstly, note that this model is still to be made
publicly available. Secondly, this incomplete model, could result in underestimation of
the mooring forces. An extensive validation research, as suggested in the next section,
for future work, can help identify the need for an improved mooring line model.
• Coupling with other models: The present research focussed on applying the CFD model
for the investigation of wave loading on floating structures. When it comes to moored
floating structures, the evaluation of the anchor strengths is an important topic. For
such an analysis it would be interesting to provide real time information of the loads in
the anchor points to a geotechnical model such as PLAXIS. As the Navier-Stokes/VOF
solver is already equipped to provide these loads, only the coupling between these
models has to be established. One can think of numerous other fields of application,
with respect to offshore wind and more generally hydrodynamical engineering, that
could benefit from highly detailed CFD computations. An example would be the instal-
lation of a gravity-based foundation of an offshore wind turbine on the seabed, again,
the coupling with a detailed geotechnical model could provide detailed information
about soil pressures and deformations through the simulation of the the touchdown
moment in the installation process.

7.4. F UTURE RESEARCH


The presented validation of the Navier-Stokes/VOF solver, with respect to wave propagation,
is considered extensive. However, the focus of the present work was on the validation of a fully
nonlinear numerical wave tank and still many research topics have been left untouched. The
two numerical models have, even without the suggested improvements, the potential to sig-
nificantly improve the computations of fluid-structure interaction in numerous applications.
For future work the following research topics could be of interest:

• Wave loads on a moored floating structure: Regretfully, the proof-of-concept compu-


tation, of the loading on a moored floating wind turbine by irregular waves, could not
be validated with experiment measurements, due limitation of time and availability of
data. This being said, the presented model can be used for the numerous subjects that
need substantial validation:
• Motions of the floating structure as a result of wave loading.
• Impact loads on the floating structure from of irregular phase-focussed waves.
• Loads in the mooring system as a result of wave impacts.
• Wave loads on secondary structures from irregular waves:
• Multi-directional phase-focused waves: Multi-directional phase-focused waves show
the potential to introduce larger peak forces than unidirectional waves. Since the work
110 7. C ONCLUSIONS AND RECOMMENDATIONS

of (Paulsen, 2013) and (Paulsen et al., 2014b) has demonstrated the excellent capabili-
ties of the potential flow solver to generate these types of free surface waves, it would be
a logical next step to investigate the effects of such waves on floating structures. Please
note that many of the research topics related to wave loads on moored floating struc-
tures, as mentioned in the previous bullet point, are interesting fields of study when it
comes to multi-directional phase-focused waves.
• Related offshore research topics: It can be argued that wave loading on floating wind
turbines is just one of the many interesting research topics regarding the search for
a cleaner offshore energy source. At the moment, many attempts are made with re-
spect to the advancements of these technologies and one could imagine using the, in
this thesis discussed, Navier-Stokes/VOF solver for the investigation of extreme wave
loads on wave energy converters, detailed modelling of tidal energy conversion sys-
tems subjected to waves, detailed towing operations and installation of gravity-based
foundations for offshore wind turbines to name a few.
A
WAVE F OAM CODE

1 / *−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−* \
========= |
3 \\ / F ield | OpenFOAM: The Open Source CFD Toolbox
\\ / O peration |
5 \\ / A nd | Copyright (C) 2011−2013 OpenFOAM Foundation
\\/ M anipulation |
7 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
License
9 This f i l e i s part of OpenFOAM.

11 OpenFOAM i s f r e e software : you can r e d i s t r i b u t e i t and/ or modify i t


under the terms of the GNU General Public License as published by
13 the Free Software Foundation , e i t h e r version 3 of the License , or
( at your option ) any l a t e r version .
15
OpenFOAM i s d i s t r i b u t e d in the hope that i t w i l l be useful , but WITHOUT
17 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 f o r more d e t a i l s .

21 You should have received a copy of the GNU General Public License
along with OpenFOAM. I f not , see <http : / /www. gnu . org / l i c e n s e s / >.
23
Application
25 interFoam

27 Description
Solver f o r 2 incompressible , isothermal immiscible f l u i d s using a VOF
29 ( volume of f l u i d ) phase−f r a c t i o n based i n t e r f a c e capturing approach .

31 The momentum and other f l u i d properties are of the " mixture " and a s i n g l e
momentum equation i s solved .
33
Turbulence modelling i s generic , i . e . laminar , RAS or LES may be selected .
35
For a two−f l u i d approach see twoPhaseEulerFoam .
37
\ *−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−* /
39
#include "fvCFD .H"

111
112 A. WAVE F OAM CODE

41 #include "CMULES.H"
#include " subCycle .H"
43 #include " i n t e r f a c e P r o p e r t i e s .H"
#include " incompressibleTwoPhaseMixture .H"
45 #include " turbulenceModel .H"
#include " pimpleControl .H"
47 #include " fvIOoptionList .H"
#include " fixedFluxPressureFvPatchScalarField .H"
49
#include " relaxationZone .H"
51 #include " externalWaveForcing .H"

53 // * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //

55 i n t main( i n t argc , char * argv [ ] )


{
57 #include " setRootCase .H"
#include " createTime .H"
59 #include "createMesh .H"

61 pimpleControl pimple (mesh) ;

63 #include " i n i t C o n t i n u i t y E r r s .H"


#include " readGravi tatio nal Accel eration .H"
65 #include " readWaveProperties .H"
#include " createExternalWaveForcing .H"
67 #include " c r e a t e F i e l d s .H"
#include " readTimeControls .H"
69 #include " createPrghCorrTypes .H"
#include " correctPhi .H"
71 #include "CourantNo .H"
#include " s e t I n i t i a l D e l t a T .H"
73
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //
75
Info << " \ nStarting time loop \n" << endl ;
77
while ( runTime . run ( ) )
79 {
#include " readTimeControls .H"
81 #include "CourantNo .H"
#include "alphaCourantNo .H"
83 #include " setDeltaT .H"

85 runTime++;

87 Info << "Time = " << runTime . timeName ( ) << nl << endl ;

89 externalWave−>step ( ) ;

91 // −−− Pressure−v e l o c i t y PIMPLE corrector loop


while ( pimple . loop ( ) )
93 {
#include " alphaControls .H"
95
i f ( pimple . f i r s t I t e r ( ) | | alphaOuterCorrectors )
97 {
113

twoPhaseProperties . c o r r e c t ( ) ;
99
#include "alphaEqnSubCycle .H"
101 relaxing . correct ( ) ;
interface . correct ( ) ;
103 }

105 #include "UEqn.H"

107 // −−− Pressure corrector loop


while ( pimple . c o r r e c t ( ) )
109 {
#include "pEqn .H"
111 }

113 i f ( pimple . turbCorr ( ) )


{
115 turbulence−>c o r r e c t ( ) ;
}
117 }

119 runTime . write ( ) ;

121 Info << "ExecutionTime = " << runTime . elapsedCpuTime ( ) << " s "
<< " ClockTime = " << runTime . elapsedClockTime ( ) << " s "
123 << nl << endl ;
}
125
// Close down the e x t e r n a l wave f o r ci ng in a nice manner
127 externalWave−>close ( ) ;

129 Info << "End\n" << endl ;

131 return 0 ;
}
133

135 // * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //
B
WAVE DY MF OAM CODE

1 / *−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−* \
========= |
3 \\ / F ield | OpenFOAM: The Open Source CFD Toolbox
\\ / O peration |
5 \\ / A nd | Copyright (C) 2011−2014 OpenFOAM Foundation
\\/ M anipulation |
7 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
License
9 This f i l e i s part of OpenFOAM.

11 OpenFOAM i s f r e e software : you can r e d i s t r i b u t e i t and/ or modify i t


under the terms of the GNU General Public License as published by
13 the Free Software Foundation , e i t h e r version 3 of the License , or
( at your option ) any l a t e r version .
15
OpenFOAM i s d i s t r i b u t e d in the hope that i t w i l l be useful , but WITHOUT
17 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 f o r more d e t a i l s .

21 You should have received a copy of the GNU General Public License
along with OpenFOAM. I f not , see <http : / /www. gnu . org / l i c e n s e s / >.
23
Application
25 interDyMFoam

27 Description
Solver f o r 2 incompressible , isothermal immiscible f l u i d s using a VOF
29 ( volume of f l u i d ) phase−f r a c t i o n based i n t e r f a c e capturing approach ,
with optional mesh motion and mesh topology changes including adaptive
31 re−meshing .

33 \ *−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−* /

35 #include "fvCFD .H"


#include "dynamicFvMesh .H"
37 #include "CMULES.H"
#include " subCycle .H"
39 #include " immiscibleIncompressibleTwoPhaseMixture .H"
#include " turbulenceModel .H"

115
116 B. WAVE DY MF OAM CODE

41 #include " pimpleControl .H"


#include " fvIOoptionList .H"
43 #include " fixedFluxPressureFvPatchScalarField .H"

45 // * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //

47 i n t main( i n t argc , char * argv [ ] )


{
49 #include " setRootCase .H"
#include " createTime .H"
51 #include "createDynamicFvMesh .H"
#include " i n i t C o n t i n u i t y E r r s .H"
53
pimpleControl pimple (mesh) ;
55
#include " c r e a t e F i e l d s .H"
57 #include " readTimeControls .H"
#include " createPrghCorrTypes .H"
59
v o l S c a l a r F i e l d rAU
61 (
IOobject
63 (
"rAU" ,
65 runTime . timeName ( ) ,
mesh,
67 IOobject : : READ_IF_PRESENT ,
IOobject : : AUTO_WRITE
69 ),
mesh,
71 dimensionedScalar ( "rAUf" , dimTime/rho . dimensions ( ) , 1 . 0 )
);
73
#include " correctPhi .H"
75 #include " createUf .H"
#include "CourantNo .H"
77 #include " s e t I n i t i a l D e l t a T .H"

79 // * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //
Info << " \ nStarting time loop \n" << endl ;
81
while ( runTime . run ( ) )
83 {
#include " readControls .H"
85 #include "alphaCourantNo .H"
#include "CourantNo .H"
87
#include " setDeltaT .H"
89
runTime++;
91
Info << "Time = " << runTime . timeName ( ) << nl << endl ;
93
// −−− Pressure−v e l o c i t y PIMPLE corrector loop
95 while ( pimple . loop ( ) )
{
97 i f ( pimple . f i r s t I t e r ( ) | | moveMeshOuterCorrectors )
117

{
99 s c a l a r timeBeforeMeshUpdate = runTime . elapsedCpuTime ( ) ;

101 mesh . update ( ) ;

103 i f (mesh . changing ( ) )


{
105 Info << " Execution time f o r mesh . update ( ) = "
<< runTime . elapsedCpuTime ( ) − timeBeforeMeshUpdate
107 << " s " << endl ;

109 gh = g & mesh .C( ) ;


ghf = g & mesh . Cf ( ) ;
111 }

113 i f (mesh . changing ( ) && correctPhi )


{
115 // Calculate absolute f l u x from the mapped surface v e l o c i t y
phi = mesh . Sf ( ) & Uf ;
117
#include " correctPhi .H"
119
// Make the f l u x r e l a t i v e to the mesh motion
121 fv c : : makeRelative ( phi , U) ;

123 mixture . c o r r e c t ( ) ;
}
125
i f (mesh . changing ( ) && checkMeshCourantNo )
127 {
#include "meshCourantNo .H"
129 }
}
131
#include " alphaControls .H"
133 #include "alphaEqnSubCycle .H"

135 mixture . c o r r e c t ( ) ;

137 #include "UEqn.H"

139 // −−− Pressure corrector loop


while ( pimple . c o r r e c t ( ) )
141 {
#include "pEqn .H"
143 }

145 i f ( pimple . turbCorr ( ) )


{
147 turbulence−>c o r r e c t ( ) ;
}
149 }

151 runTime . write ( ) ;

153 Info << "ExecutionTime = " << runTime . elapsedCpuTime ( ) << " s "
<< " ClockTime = " << runTime . elapsedClockTime ( ) << " s "
118 B. WAVE DY MF OAM CODE

155 << nl << endl ;


}
157
Info << "End\n" << endl ;
159
return 0 ;
161 }

163
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
165 //
C
S IX D O FR IGID B ODY M OTION CODE

/ *−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−* \
2 ========= |
\\ / F ield | OpenFOAM: The Open Source CFD Toolbox
4 \\ / O peration |
\\ / A nd | Copyright (C) 2011−2015 OpenFOAM Foundation
6 \\/ M anipulation |
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
8 License
This f i l e i s part of OpenFOAM.
10 OpenFOAM i s f r e e software : you can r e d i s t r i b u t e i t and/ or modify i t
under the terms of the GNU General Public License as published by
12 the Free Software Foundation , e i t h e r version 3 of the License , or
( at your option ) any l a t e r version .
14 OpenFOAM i s d i s t r i b u t e d in the hope that i t w i l l be useful , but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
f o r more d e t a i l s .
18 You should have received a copy of the GNU General Public License
along with OpenFOAM. I f not , see <http : / /www. gnu . org / l i c e n s e s / >.
20 \ *−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−* /

22 #include "sixDoFRigidBodyMotion .H"


#include " septernion .H"
24
// * * * * * * * * * * * * * P r i v a t e Member Functions * * * * * * * * * * * //
26
void Foam : : sixDoFRigidBodyMotion : : applyRestraints ( )
28 {
i f ( r e s t r a i n t s _ . empty ( ) )
30 {
return ;
32 }

34 i f ( Pstream : : master ( ) )
{
36 forAll ( restraints_ , rI )
{
38 i f ( report_ )
{
40 Info << " R e s t r a i n t " << r e s t r a i n t s _ [ r I ] . name( ) << " : " ;

119
120 C. S IX D O FR IGID B ODY M OTION CODE

}
42
// R e s t r a i n t position
44 point rP = vector : : zero ;

46 // R e s t r a i n t force
vector rF = vector : : zero ;
48
// R e s t r a i n t moment
50 vector rM = vector : : zero ;

52 // Accumulate the r e s t r a i n t s
r e s t r a i n t s _ [ r I ] . r e s t r a i n ( * t h i s , rP , rF , rM) ;
54
// Update the a c c e l e r a t i o n
56 a ( ) += rF /mass_ ;

58 // Moments are returned in global axes , transforming to


// body l o c a l to add to torque .
60 tau ( ) += Q( ) . T ( ) & (rM + ( ( rP − centreOfRotation ( ) ) ^ rF ) ) ;
}
62 }
}
64

66 // * * * * * * * * * * * * * * * * Constructors * * * * * * * * * * * * * * //

68 Foam : : sixDoFRigidBodyMotion : : sixDoFRigidBodyMotion ( )


:
70 motionState_ ( ) ,
motionState0_ ( ) ,
72 restraints_ () ,
constraints_ ( ) ,
74 tConstraints_ ( tensor : : I ) ,
rConstraints_ ( tensor : : I ) ,
76 initialCentreOfMass_ ( vector : : zero ) ,
i n i t i a l C e n t r e O f R o t a t i o n _ ( vector : : zero ) ,
78 initialQ_ ( I ) ,
mass_ (VSMALL) ,
80 momentOfInertia_ ( diagTensor : : one *VSMALL) ,
aRelax_ ( 1 . 0 ) ,
82 aDamp_( 1 . 0 ) ,
report_ ( f a l s e )
84 {}

86
Foam : : sixDoFRigidBodyMotion : : sixDoFRigidBodyMotion
88 (
const d i cti ona ry& dict ,
90 const d i cti ona ry& s t a t e D i c t
)
92 :
motionState_ ( s t a t e D i c t ) ,
94 motionState0_ ( ) ,
restraints_ () ,
96 constraints_ ( ) ,
tConstraints_ ( tensor : : I ) ,
121

98 rConstraints_ ( tensor : : I ) ,
initialCentreOfMass_
100 (
d i c t . lookupOrDefault
102 (
" initialCentreOfMass " ,
104 vector ( d i c t . lookup ( " centreOfMass " ) )
)
106 ),
i n i t i a l C e n t r e O f R o t a t i o n _ ( initialCentreOfMass_ ) ,
108 initialQ_
(
110 d i c t . lookupOrDefault
(
112 " initialOrientation " ,
d i c t . lookupOrDefault ( " o r i e n t a t i o n " , tensor : : I )
114 )
),
116 mass_ ( readScalar ( d i c t . lookup ( "mass" ) ) ) ,
momentOfInertia_ ( d i c t . lookup ( "momentOfInertia" ) ) ,
118 aRelax_ ( d i c t . lookupOrDefault < scalar >( " accelerationRelaxation " , 1 . 0 ) ) ,
aDamp_( d i c t . lookupOrDefault < scalar >( " accelerationDamping " , 1 . 0 ) ) ,
120 report_ ( d i c t . lookupOrDefault <Switch >( " report " , f a l s e ) )
{
122 addRestraints ( d i c t ) ;

124 // Set c o n s t r a i n t s and i n i t i a l centre of r o t a t i o n


// i f d i f f e r e n t to the centre of mass
126 addConstraints ( d i c t ) ;

128 // I f the centres of mass and r o t a t i o n are d i f f e r e n t . . .


vector R( initialCentreOfMass_ − i n i t i a l C e n t r e O f R o t a t i o n _ ) ;
130 i f ( magSqr (R) > VSMALL)
{
132 // . . . c o r r e c t the moment of i n e r t i a tensor using p a r a l l e l axes theorem
momentOfInertia_ += mass_ * diag ( I * magSqr (R) − sqr (R) ) ;
134
// . . . and i f the centre of r o t a t i o n i s not s p e c i f i e d f o r motion s t a t e
136 // update i t
i f ( ! s t a t e D i c t . found ( " centreOfRotation " ) )
138 {
motionState_ . centreOfRotation ( ) = i n i t i a l C e n t r e O f R o t a t i o n _ ;
140 }
}
142
// Save the old−time motion s t a t e
144 motionState0_ = motionState_ ;
}
146

148 Foam : : sixDoFRigidBodyMotion : : sixDoFRigidBodyMotion


(
150 const sixDoFRigidBodyMotion& sDoFRBM
)
152 :
motionState_ (sDoFRBM. motionState_ ) ,
154 motionState0_ (sDoFRBM. motionState0_ ) ,
122 C. S IX D O FR IGID B ODY M OTION CODE

r e s t r a i n t s _ (sDoFRBM. r e s t r a i n t s _ ) ,
156 c o n s t r a i n t s _ (sDoFRBM. c o n s t r a i nt s _ ) ,
tConstraints_ (sDoFRBM. tConstraints_ ) ,
158 rConstraints_ (sDoFRBM. rConstraints_ ) ,
initialCentreOfMass_ (sDoFRBM. initialCentreOfMass_ ) ,
160 i n i t i a l C e n t r e O f R o t a t i o n _ (sDoFRBM. i n i t i a l C e n t r e O f R o t a t i o n _ ) ,
i n i t i a l Q _ (sDoFRBM. i n i t i a l Q _ ) ,
162 mass_ (sDoFRBM. mass_ ) ,
momentOfInertia_ (sDoFRBM. momentOfInertia_ ) ,
164 aRelax_ (sDoFRBM. aRelax_ ) ,
aDamp_(sDoFRBM.aDamp_) ,
166 report_ (sDoFRBM. report_ )
{}
168

170 // * * * * * * * * * * * * * * Member Functions * * * * * * * * * * * * * * //

172 void Foam : : sixDoFRigidBodyMotion : : addRestraints


(
174 const d i cti ona ry& d i c t
)
176 {
i f ( d i c t . found ( " r e s t r a i n t s " ) )
178 {
const d i cti onar y& r e s t r a i n t D i c t = d i c t . subDict ( " r e s t r a i n t s " ) ;
180
label i = 0;
182
restraints_ . setSize ( restraintDict . size ( ) ) ;
184
f o r A l l C o n s t I t e r ( IDLList <entry > , r e s t r a i n t D i c t , i t e r )
186 {
i f ( i t e r ( ) . isDict ( ) )
188 {
restraints_ . set
190 (
i ++ ,
192 sixDoFRigidBodyMotionRestraint : : New
(
194 i t e r ( ) . keyword ( ) ,
i t e r ( ) . dict ( )
196 )
);
198 }
}
200
restraints_ . setSize ( i ) ;
202 }
}
204

206 void Foam : : sixDoFRigidBodyMotion : : addConstraints


(
208 const d i cti ona ry& d i c t
)
210 {
i f ( d i c t . found ( " c o n s t r a i n t s " ) )
123

212 {
const d i cti onar y& constraintDict = d i c t . subDict ( " c o n s t r a i n t s " ) ;
214
label i = 0;
216
c o n s tr a i n t s _ . s e t S i z e ( constraintDict . s i z e ( ) ) ;
218
pointConstraint pct ;
220 pointConstraint pcr ;

222 f o r A l l C o n s t I t e r ( IDLList <entry > , constraintDict , i t e r )


{
224 i f ( i t e r ( ) . isDict ( ) )
{
226 constraints_ . set
(
228 i,
sixDoFRigidBodyMotionConstraint : : New
230 (
i t e r ( ) . keyword ( ) ,
232 i t e r ( ) . dict ( ) ,
* this
234 )
);
236
c o n s t r a i n t s _ [ i ] . setCentreOfRotation ( i n i t i a l C e n t r e O f R o t a t i o n _ ) ;
238 c o n s t r a i n t s _ [ i ] . constrainTranslation ( pct ) ;
c o n s t r a i n t s _ [ i ] . constrainRotation ( pcr ) ;
240
i ++;
242 }
}
244
c o n s tr a i n t s _ . s e t S i z e ( i ) ;
246
tConstraints_ = pct . constraintTransformation ( ) ;
248 rConstraints_ = pcr . constraintTransformation ( ) ;

250 Info << " T r a n s l a t i o n a l co nstrai nt tensor " << tConstraints_ << nl
<< " Rotational cons trai nt tensor " << rConstraints_ << endl ;
252 }
}
254

256 void Foam : : sixDoFRigidBodyMotion : : updatePosition


(
258 s c a l a r deltaT ,
s c a l a r deltaT0
260 )
{
262 // F i r s t leapfrog v e l o c i t y adjust and motion part , required before
// force c a l c u l a t i o n
264
i f ( Pstream : : master ( ) )
266 {
v ( ) = tConstraints_ & ( v0 ( ) + aDamp_* 0.5 * deltaT0 * a ( ) ) ;
268 pi ( ) = rConstraints_ & ( pi0 ( ) + aDamp_* 0.5 * deltaT0 * tau ( ) ) ;
124 C. S IX D O FR IGID B ODY M OTION CODE

270 // Leapfrog move part


centreOfRotation ( ) = centreOfRotation0 ( ) + deltaT * v ( ) ;
272
// Leapfrog o r i e n t a t i o n adjustment
274 Tuple2<tensor , vector > Qpi = r o t a t e (Q0 ( ) , pi ( ) , deltaT ) ;
Q( ) = Qpi . f i r s t ( ) ;
276 pi ( ) = rConstraints_ & Qpi . second ( ) ;
}
278
Pstream : : s c a t t e r ( motionState_ ) ;
280 }

282
void Foam : : sixDoFRigidBodyMotion : : updateAcceleration
284 (
const vector& fGlobal ,
286 const vector& tauGlobal ,
s c a l a r deltaT
288 )
{
290 s t a t i c bool f i r s t = f a l s e ;

292 // Second leapfrog v e l o c i t y adjust part , required a f t e r motion and


// a c c e l e r a t i o n c a l c u l a t i o n
294
i f ( Pstream : : master ( ) )
296 {
// Save the previous i t e r a t i o n a c c e l e r a t i o n s f o r r e l a x a t i o n
298 vector a P r e v I t e r = a ( ) ;
vector t a u P r e v I t e r = tau ( ) ;
300
// Calculate new a c c e l e r a t i o n s
302 a ( ) = fGlobal /mass_ ;
tau ( ) = (Q( ) . T ( ) & tauGlobal ) ;
304 applyRestraints ( ) ;

306 // Relax a c c e l e r a t i o n s on a l l but f i r s t i t e r a t i o n


if (! first )
308 {
a ( ) = aRelax_ * a ( ) + (1 − aRelax_ ) * a P r e v I t e r ;
310 tau ( ) = aRelax_ * tau ( ) + (1 − aRelax_ ) * t a u P r e v I t e r ;
}
312 f i r s t = false ;

314 // Correct v e l o c i t i e s
v ( ) += tConstraints_ & aDamp_* 0.5 * deltaT * a ( ) ;
316 pi ( ) += rConstraints_ & aDamp_* 0.5 * deltaT * tau ( ) ;

318 i f ( report_ )
{
320 status ( ) ;
}
322 }

324 Pstream : : s c a t t e r ( motionState_ ) ;


}
125

326

328 void Foam : : sixDoFRigidBodyMotion : : s t a t u s ( ) const


{
330 Info << "6−DoF r i g i d body motion" << nl
<< " Centre of r o t a t i o n : " << centreOfRotation ( ) << nl
332 << " Centre of mass : " << centreOfMass ( ) << nl
<< " Orientation : " << o r i e n t a t i o n ( ) << nl
334 << " Linear v e l o c i t y : " << v ( ) << nl
<< " Angular v e l o c i t y : " << omega ( )
336 << endl ;
}
338

340 Foam : : tmp<Foam : : pointField > Foam : : sixDoFRigidBodyMotion : : transform


(
342 const pointField& i n i t i a l P o i n t s
) const
344 {
return
346 (
centreOfRotation ( )
348 + (Q( ) & i n i t i a l Q _ . T ( ) & ( i n i t i a l P o i n t s − i n i t i a l C e n t r e O f R o t a t i o n _ ) )
);
350 }

352
Foam : : tmp<Foam : : pointField > Foam : : sixDoFRigidBodyMotion : : transform
354 (
const pointField& i n i t i a l P o i n t s ,
356 const s c a l a r F i e l d& s c a l e
) const
358 {
// Calculate the transformation septerion from the i n i t i a l s t a t e
360 septernion s
(
362 centreOfRotation ( ) − i n i t i a l C e n t r e O f R o t a t i o n ( ) ,
quaternion (Q( ) & i n i t i a l Q ( ) . T ( ) )
364 );

366 tmp<pointField > tpoi nts (new pointField ( i n i t i a l P o i n t s ) ) ;


pointField& points = tpoi nts ( ) ;
368
f o r A l l ( points , p o i n t i )
370 {
// Move non−s t a t i o n a r y points
372 i f ( s c a l e [ p o i n t i ] > SMALL)
{
374 // Use solid −body motion where s c a l e = 1
i f ( s c a l e [ p o i n t i ] > 1 − SMALL)
376 {
points [ p o i n t i ] = transform ( i n i t i a l P o i n t s [ p o i n t i ] ) ;
378 }
// Slerp septernion i n t e r p o l a t i o n
380 else
{
382 septernion ss ( s l e r p ( septernion : : I , s , s c a l e [ p o i n t i ] ) ) ;
126 C. S IX D O FR IGID B ODY M OTION CODE

384 points [ p o i n t i ] =
initialCentreOfRotation ( )
386 + ss . transform
(
388 i n i t i a l P o i n t s [ pointi ]
− initialCentreOfRotation ( )
390 );
}
392 }
}
394
return t poi nts ;
396 }

398
// * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * //
D
F LOATING WIND TURBINE MODEL

Figure D.1: Two-dimensional representation of the floating wind turbine model.

127
128 D. F LOATING WIND TURBINE MODEL

Figure D.2: Two-dimensional representation of the sub-structure of the floating wind turbine model, where the
Froude scale was 1:50. (a) left side view, (b) back view. The dimensions are: 1 = 0.06m, 2 = 0.52m, 3 = 0.817m, 4 =
0.577m, 5 = 0.50m, 6 = 0.817m, 7 = 0.48m, 8 = 0.12m, 9 = 1.48m and 10 = 1.00m.

Table D.1: Coordinates of COG and restraining system of the floating wind turbine model depicted in Figure D.1
and Figure D.2, where the Froude scale was 1:50.
Indication Description Attachment Anchor
[m] [m]
I COG {0.0, 0.0, 0.1186} -
II front mooring {0.818, 0.0, 0.0} {18.434, 0.0, 0.0}
III back mooring {-0.818, 0.0, 0.0} {-16.762, 0.0, 0.0}
IV left spring {-0.408, -0.620, 0.48} {-0.408, -2.0, 0.48}
V right spring {-0.408, 0.620, 0.48} {-0.408, 2.0, 0.48}
129

Table D.2: Dimensions of the sub-structure of the floating wind turbine model depicted in Figure D.2, where the
Froude scale was 1:50.
Number Diameter length
Indication Description
of pieces [m] [m]
A transition 1 0.14 0.12
B center column 1 0.13 0.52
C offset columns 3 0.24 0.52
D heave plates 3 0.48 0.12
E braces 9 0.032 -

Table D.3: Properties of the restraining system of the floating wind turbine model depicted in Figure D.1 and
Figure D.2. Wsub is the submerged weight of the mooring line and k spr i ng is the spring stiffness, where the
Froude scale was 1:50.
Description Length (at rest) M sub / k spr i ng
[m] [N/m]
Front mooring 18.38 0.04348
Back mooring 16.71 0.04348
Left spring 1.2114 1.51886
Right spring 1.2114 1.51886

Table D.4: Properties of the floating wind turbine model depicted in Figure D.1 and Figure D.2. These are proper-
ties including the floater, turbine and tower.
Description Symbol Unit Magnitude
Draft (mooring) T [m] 0.40
Draft (no mooring) T [m] 0.3898
Mass M [kg] 111.664
Roll inertia I xx [kg/m2 ] 49.768
Pitch inertia Iyy [kg/m2 ] 47.556
Yaw inertia I zz [kg/m2 ] 43.814
130 D. F LOATING WIND TURBINE MODEL

Figure D.3: Detailed two-dimensional representation of the computational mesh of the sub-structure of the float-
ing wind turbine model.
B IBLIOGRAPHY
Arapogianni, A., Genachte, A., Ochagavia, R., Vergara, J., Castell, D., Tsouroukdissian, A., Ro-
driguez, J., Korbijn, J., Bolleman, N., Huera-Huarte, F., and Schuon, F. (2013). Deep water:
the next step for offshore wind energy. Technical report, European Wind Energy Associa-
tion (EWEA), Brussels.
Arapogianni, A., Moccia, J., Williams, D., Phillips, J., and Hassan, G. G. (2011). Wind in our
sails: The coming of Europe’s offshore wind energy industry. Technical report, The Euro-
pean Wind Energy Association, Brussels.
Berberović, E., Van Hinsberg, N., Jakirlić, S., Roisman, I., and Tropea, C. (2009). Drop impact
onto a liquid layer of finite thickness: Dynamics of the cavity evolution. Physical Review E,
79(3):1–15.
Bos, F. M., van Oudheusden, B. W., and Bijl, H. (2013). Radial basis function based mesh
deformation applied to simulation of flow around flapping wings. Computers & Fluids,
79:167–177.
Dam, E. B., Koch, M., and Lillholm, M. (1998). Quaternions , interpolation and animation.
Datalogisk Institut, Københavns Universitet.
Delhommeau, G. (1993). Seakeeping Codes AQUADYN and AQUAPLUS. 19th WEGEMT
School Numerical Simulation of Hydrodynamics: Ships and Offshore Structures.
Det Norske Veritas (2014). DNV-OS-J101 offshore standard. Design of offshore wind turbine
structures.
Dixon, A., Greated, C. A., and Salter, S. H. (1979). Wave forces on partially submerged cylin-
ders. Journal of the Waterway Port Coastal and Ocean Division, 105(4):421–438.
Dunbar, A. J., Craven, B. a., and Paterson, E. G. (2015). Development and validation of a tightly
coupled CFD/6-DOF solver for simulating floating offshore wind turbine platforms. Ocean
Engineering, 110:98–105.
Engsig-Karup, A. P., Bingham, H. B., and Lindberg, O. (2009). An efficient flexible-order model
for 3D nonlinear water waves. Journal of Computational Physics, 228(6):2100–2118.
Faltinsen, O. M., Newman, J. N., and Vinje, T. (1995). Nonlinear wave loads on a slender
vertical cylinder. Journal of Fluid Mechanics, 289:179–198.
Fenton, J. D. (1988). The numerical solution of steady water wave problems. Computers &
Geosciences, 14(3):357–368.
Fenton, J. D. (1990). Nonlinear wave theories. The Sea: Ocean Engineering Science, 9(1):3–25.
Ferziger, J. and Peric, M. (2012). Computational methods for fluid dynamics. Springer Science
& Business Media.
Forristall, G. Z. (1978). On the statistical distribution of wave heights in a storm. Journal of
Geophysical Research, 83(5):2353–2358.
Gopala, V. R. and van Wachem, B. G. M. (2008). Volume of fluid methods for immiscible-fluid
and free-surface flows. Chemical Engineering Journal, 141(1):204–221.
Grue, J. and Huseby, M. (2002). Higher harmonic wave forces and ringing of vertical cylinders.
Applied Ocean Research, 24(4):203–214.
Hamilton, W. R. (1844). On quaternions, or on a new system of imaginaries in alge-
bra. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science,
25(163):10–13.
Hirt, C. and Nichols, B. (1981). Volume of fluid (VOF) method for the dynamics of free bound-
aries. Journal of Computational Physics, 39(1):201–225.

131
132 B IBLIOGRAPHY

Ito, S. (1977). Study of the transient heave oscillation of a floating cylinder. PhD thesis, Mas-
sachusetts Institute of Technology.
Jacobsen, N. G., Fuhrman, D. R., and Fredsøe, J. (2012). A wave generation toolbox for the
open-source CFD library: OpenFoam®. International Journal for Numerical Methods in
Fluids, 70(9):1073–1088.
Jonkman, J., Butterfield, S., Musial, W., and Scott, G. (2009). Definition of a 5-MW reference
wind turbine for offshore system development. Technical report, National Renewable En-
ergy Laboratory Golden, USA.
Jonkman, J. and Musial, W. (2010). Offshore code comparison collaboration (OC3) for IEA
task 23 offshore wind technology and deployment. Contract, 303(December):1–70.
Jonkman, J., Robertson, A. N., Popko, W., Vorpahl, F., Zuga, A., Kohlmeier, M., Larsen, T. J.,
Yde, A., Saetertro, K., Okstad, K. M., Nichols, J., Nygaard, T. a., Gao, Z., Manolas, D., Kim, K.,
Yu, Q., Shi, W., Park, H., Vasquez-Rojas, A., Dubois, J., Kaufer, D., Thomassen, P., de Ruiter,
M. J., Peeringa, J., Zhiwen, H., and von Waaden, H. (2012). Offshore Code Comparison
Collaboration Continuation (OC4), Phase I: results of coupled simulations of an offshore
wind turbine with jacket support structure. International Society of Offshore and Polar
Engineers, 1:1–11.
Koo, B. J., Goupee, A. J., Kimball, R. W., and Lambrakos, K. F. (2014). Model tests for a floating
wind turbine on three different floaters. Offshore Mechanics and Arctic Engineering, 136(2).
Krenk, S. (2001). Mechanics and analysis of beams, columns and cables: a modern introduc-
tion to the classic theories. Springer Science & Business Media.
Kring, D. C., Korsmeyer, F. T., Singer, J., Danmeier, D., and White, J. (1999). Accelerated non-
linear wave simulations for large structures. 7th International Conference on Numerical
Ship Hydrodynamics, Nantes, France.
Lee, C. (1995). WAMIT theory manual. Department of Ocean engineering, Massachusetts
Institute of Technology.
Madsen, P. A. and Fuhrman, D. R. (2010). High-order Boussinesq-type modeling of nonlinear
wave phenomena in deep and shallow water. Advances in numerical simulation of nonlin-
ear water waves, 5:245–285.
Malenica, Š. and Molin, B. (1995). Third-harmonic wave diffraction by a vertical cylinder.
Journal of Fluid Mechanics, 302:203–229.
Maruyama, D., Bailly, D., and Carrier, G. (2012). High quality mesh deformation using quater-
nions for orthogonality preservation. American Institute of Aeronautics and Astronautics,
52(12):9–12.
Maskell, S. J. and Ursell, F. (1970). The transient motion of a floating body. Journal of Fluid
Mechanics, 44(2):303–313.
Matha, D., Schlipf, M., Cordle, A., Pereira, R., and Jonkman, J. (2011). Challenges in simulation
of aerodynamics , hydrodynamics , and mooring-line dynamics of floating offshore wind
turbines. International Offshore and Polar Engineering, 8:421–428.
Mayer, S., Garapon, A., and Sørensen, L. (1998). A fractional step method for unsteady free
surface flow with applications to nonlinear wave dynamics. International Journal for Nu-
merical Methods in Fluids, 28(2):293–315.
Moccia, J., Wilkes, J., Pineda, I., and Corbetta, G. (2014). Wind energy scenarios for 2020.
Technical report, The European Wind Energy Association, Brussels.
Morison, J. R., Johnson, J. W., and Schaaf, S. A. (1950). The force exerted by surface waves on
piles. Journal of Petroleum Technology, 2(5):149–154.
Omidvar, P. (2010). Wave loading on bodies in the free surface using Smoothed Particle Hydro-
dynamics (SPH). PhD thesis, The University of Manchester, UK.
Osher, S. and Sethian, J. a. (1988). Fronts propagating with curvature dependent speed:
algorithms based on Hamilton-Jacobi formulations. Journal of Computational Physics,
B IBLIOGRAPHY 133

79(1):12–49.
Paulsen, B. T. (2013). Efficient computations of wave loads on off shore structures. PhD thesis,
Technical University of Denmark.
Paulsen, B. T., Bredmose, H., Bingham, H., and Jacobsen, N. (2014a). Forcing of a bottom-
mounted circular cylinder by steep regular water waves at finite depth. Journal of Fluid
Mechanics, 755:1–34.
Paulsen, B. T., Bredmose, H., and Bingham, H. B. (2014b). An efficient domain decomposition
strategy for wave loads on surface piercing circular cylinders. Coastal Engineering, 86:57–
76.
Pineda, I., Azau, S., Moccia, J., and Wilkes, J. (2014). Wind in power: 2013 european statistics.
Technical report, The European Wind Energy Association, Brussels.
Rainey, R. C. T. (2007). Weak or strong nonlinearity: the vital issue. Journal of Engineering
Mathematics, 58(1-4):229–249.
Rienecker, M. M. and Fenton, J. D. (1981). A Fourier approximation method for steady water
waves. Journal of Fluid Mechanics, 104:119–137.
Samareh, J. A. (2002). Application deformation of quaternions for mesh deformation. Tech-
nical report, Langley Research Center, Hampton, Virginia.
Sarpkaya, T. and Isaacson., M. (1981). Mechanics of wave forces on offshore structures. Van
Nostrannd Reinhold Company Inc.
Schäffer, H. A. and Steenberg, C. M. (2003). Second-order wavemaker theory for multidirec-
tional waves. Ocean Engineering2, 30(10):1203–1231.
Seng, S. (2012). Slamming and whipping analysis of ships. PhD thesis, Technical University
of Denmark.
Shoemake, K. (1985). Animating rotation with quaternion curves. Computer Graphics,
19(3):245–254.
Waisman, F., Gurley, K., Grigoriu, M., and Kareem, A. (2002). Non-Gaussian model for ringing
phenomena in offshore structures. Journal of Engineering Mechanics, 128(7):730–741.
Wellens, R. (2012). Wave simulation in truncated domains for offshore applications. PhD
thesis, Delft University of Technology, The Netherlands.
Williams, J. M. (1981). Limiting gravity waves in water of finite depth. Philosophical
Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences,
302(1466):139–188.
Yu, Y. H. and Li, Y. (2013). Reynolds-Averaged Navier-Stokes simulation of the heave perfor-
mance of a two-body floating-point absorber wave energy system. Computers & Fluids,
73:104–114.
Yu, Y. S. and Ursell, F. (1961). Surface waves generated by an oscillating circular cylinder on
water of finite depth: theory and experiment. Journal of Fluid Mechanics, 11(4):529–551.

Das könnte Ihnen auch gefallen